https://kangax.substack.com/feed

CrossFit apps are good at storing WOD scores, but are terrible at knowing the work you’ve actually done. Here’s a WOD from my box:

Out of 20 results, only 1 person did it RX:

In SugarWOD the only way to mark WOD as scaled is via notes. Notice how everyone wrote 155#, 145 lb, 165lb, @165, etc.

Ok, so what’s the problem?

Well, the reality is that I’ve done 10 sets of 2 power cleans + 1 power jerk @165lb. My training log in SugarWOD, however, has no idea that this is what I’ve done. If I go into “History” for this WOD and click on Apr 24 I see “170” which makes me think that maybe that’s the previous weight I did. But no, it brings me to a WOD where I did “1 above knee squat snatch @170lb” not “2 Power Clean + 1 Split Jerk” or even just “Power Clean + Split Jerk” or even just a standalone “Power Clean” / “Split Jerk”:

It’s entirely wrong and confusing. And if I go into my barbell lift history, this weight I just lifted is neither under Power Clean, nor under Clean and Jerk, nor under Split Jerk:

The two lonely logs for Power Clean are from Mar 2024. That’s… 2 years ago. SugarWOD has no idea about countless times I’ve done power clean in a WOD like today.

My barbell PR history in an app simply can’t be trusted.

Ok, that’s SugarWOD. What about others? Of ~25 apps, only BTWB and WodUp treat scaled work as structured data — and both punt on the hard part. BTWB excludes scaled metcons from its fitness rating. WodUp tags them four ways but disqualifies any scaled benchmark from PR tracking. Everyone else — Wodify, SugarWOD, Mayhem, CompTrain, TrainHeroic, Hevy, Strong, the long tail — collapses your work into Rx/Scaled + a score + a notes field.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

In PRzilla, we already tracked all movements correctly when logged individually or as part of a WOD. Here’s a workout with Power Clean that I did in Apr 2025 which SugarWOD never showed. It’s displayed in the history, alongside individual lifts:

Scaled WODs like this Holleyman were part of the history too, but we didn’t actually know how you scaled them outside of the notes. It would just display as "(scaled)” and not be part of any metrics:

I always knew this is one of the most important things to have in a CrossFit app worth its salt; after many weeks of work, it’s finally here.

You can now specify exactly the work you’ve done when logging any WOD in PRzilla. WOD’s are just prescriptions. Every WOD is decomposed into movement blocks. And it’s those movement blocks that define your history, stats, charts, etc. in our system.

But here’s a harsh truth: the last thing athletes want to do when logging WODs is to spend few minutes tinkering with movements, sets, reps, and weights. Sweaty hands. Still breathing hard. 99% of us write “165#” and move on. That’s why it was very important for me to remove as much friction when logging scaled WODs as possible.

Let’s go back to the Clean & Jerk WOD. The moment I add it, PRzilla parses it into Power Clean, Split Jerk, and Air Bike; it creates all 10 sets of each, prefills first with 2 reps, second with 1 rep, and adds RX weight to both.

Because I scaled to 165lb, all I need to do is type “165” in Power Clean’s topmost weight cell and click “Apply to all sets below”. It takes me the same amount of time as writing “165#”.

If I scaled that WOD to just power clean I can remove Split Jerk movement from the session by clicking “Delete”. It’s actually faster than typing “Did not do Split Jerk” in the notes. And I can replace power clean with dumbbell clean just as quickly as typing “Scaled to db cleans”.

We now have the power of a full-blown workout editor within the context of one WOD. And we have minimal friction.

When looking at your workout session (left pic), you now see the things you’ve actually done and not just what was prescribed. A lot more useful metric that goes in line with the rest of the movements in a workout session. After all, we show 3 x 2 @ 275lb on Front Squats and not just “Find 2RM in 15min”:

The Holleyman (right pic) is now displaying 30 x 1 @155lb because we have your actual weight as part of the WOD movements (=work performed). Seeing all the weights in a history list like this feels like a basic right in a CrossFit-aware workout tracker :)

Ok, charts and history lists are nice and all but… does this all really matter? I think so. Accurate training data is the foundation of your fitness training. We need to know where you stand in order to design a path forward.

If you scaled Holleyman to 185lb (and finished all 30 rounds), now we know these things:

1. 1RM — this tells us that your max is at least 185lb without explicitly testing it.
   1RM is essential for daily programming.

2. Volume — you’ve done 185x30=5550lb of Power Clean in that session.
   When we know your volume, we know the intensity of a given workout relative to your usual effort. If you’ve only ever done 1000lb of power cleans in a WOD, doing a WOD with 5000lb will wreck you for days. We can mark such WOD as high-volume for your level. A high-volume day can then either be scaled correctly or programmed in a way that doesn’t interfere with other workouts in a week.

3. Progress — Holleyman @155lb in 20min vs. Holleyman @185lb in 20min is the kind of progress you can see if we track scaled work. That’s density: more work done in the same amount of time. It’s important to know if you’re trending up, down or plateauing.

4. Programming — passing your results to AI coach makes for highly relevant advice. If it knows you did 30 rounds of power cleans @185lb, it can use that to create a workout that’s within your reach. It can tell you to scale up or down next time according to this weight. It can even flag less obvious things like weak squat clean in relation to power clean.

A coach who doesn't know what you lifted can't tell you what to lift tomorrow. Neither can your app. PRzilla is my attempt to fix that — give it a try.

If you’ve been following this blog, you know I’ve spent the past year building PRzilla — from AI-powered WOD benchmarks to My Fitness visualizations that replaced a decade of spreadsheets to a full workout tracker. If you’re new here, the short version: I left my tech career to build the CrossFit training app I always wanted, and it turned into something much bigger than I expected.

The native app is live — iOS and Android, 11 consecutive weekly releases and counting — and what started as a workout tracker has become an AI training companion that sits on top of your data, reads your recovery metrics, and coaches you like someone who actually knows your training history.

Here’s how I got there.

As I mentioned earlier, most training apps I’ve used are either geared towards traditional strength training (Strong/Hevy) or are Crossfit-class-centric and suck at logging any other activity (SugarWOD/Wodify/PushPress) or are too specialized and have poor workout trackers (Strava focuses on running/biking and Whoop focuses on recovery; strength logging is poor).

If I joined my gym class on Monday, I want to log whatever WOD was on a whiteboard / in SugarWOD. If on Tuesday I do an open gym instead to work on my ring muscle up practice, I want to log 3-3-3-2-2-2-2 and let it be part of my metrics. Maybe I also do an hour on a rower to work on zone 2. Or maybe I went for a run with a friend who's preparing for Hyrox or a Brooklyn half. I want to log those too. And then perhaps on Friday I'm traveling to SF and will drop into another box. The box is on a completely different programming track and often different platform like Wodify or PushPress. I want to log whatever WOD I end up doing at that gym. And even if I only ever do my gym programming via SugarWOD, what happens when I move to another area and join another box that uses different software—I’d like to keep all my history and lifts.

PRzilla solves all of this. Here's what that looks like in practice:

1. No longer tied to a single CrossFit box and their platform

2. You can log anything you want

3. You own your data

One feature I’m especially excited about is that you can snap a photo of your gym’s whiteboard and the app breaks the entire training day into structured movements and WODs — matching canonical benchmarks when it recognizes them or creating a custom one when it doesn’t. WODs automatically get analyzed for performance level bell curve so you can immediately see where you stand.

Here I snapped SugarWOD’s whiteboard and it turned it into back squat and shoulder press as single movements (which I could then log as many sets/reps of) and it created a WOD with running and Fran-like couplet, giving me a time to shoot for:

I spent months building and refining this sophisticated workout tracker that can handle any workout. Even though I made it easy to log things, I eventually realized that training log itself is not the end goal. There are two more legs of a tripod that truly elevate training experience to the next level.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Your training log is the past. Your programming is the future.

They’re intimately related to each other: as a coach, I need to know what you’re capable of doing in order to design a program to get you to your goal. Your training log is a snapshot of your abilities, strengths and weaknesses. An advanced lifter that needs to work on their positions in a snatch to bring it from 1.25x to 1.5xBW has very different movements prescribed than someone who just started training and are working on their 5x5 Back Squat progression.

However, Training Data ↔ Programming don’t exist in a vacuum; the subject in the middle—you—responds to demands placed on them (=programming) positively or negatively. One way to measure how they respond to them is via your recovery metrics: HRV, RHR, Sleep hours/quality. Waking up with low HRV could be related to a hard session yesterday (it tells us that it might have been too intense for you). And it should adjust your programming accordingly (multiple hard sessions lead to overtraining).

And so I’ve added Apple Health integration:

The feature is in beta but the idea is that y

ou can now see your “readiness” day-by-day. Last week I saw my HRV drop 15ms after back-to-back heavy sessions — a clear signal I was accumulating more fatigue than I was recovering from. Without this data, I would have pushed through another hard day. With it, I dialed back to skill work and was fully recovered by Thursday.

A good coach would start by looking at your training history, create a program to get you from point A (your past) to point B (your goal), then adjust it in real time based on how you respond to it (your recovery metrics).

Given AI capabilities in 2026, it can do all of the above. It just needs to have access to your health and training data. It can serve as an orchestrator in the middle, carefully analyzing your past (recent training), ingesting what’s present (morning readiness) and adjusting your future (today or tomorrow’s workout).

Meet your personal coach; they know you better than a coach in your gym.

The coach has access to your full training history, remembers past conversations, and can search the web for workout details you mention. You don't re-explain yourself every session — it already knows your recent lifts, your weak points, and your goals.

As you can see, I’ve been using it for my own training. My honest take, as a coach myself, is that I’m genuinely surprised at how relevant the answers are. I would say they’re at ~80-90% of what a human coach would recommend; not perfect and sometimes sloppy, but very useful as a brainstorming partner if you’re a high level athlete, and are very illuminating if you’re a beginner.

Some things you can ask:

- “How has my squat been progressing?”
- “How should I approach this workout?” (attach a WOD from our catalog or a picture)
- “What movements have I not done much lately that decay the fastest?”
- “Help me structure this week as a best prep for QuarterFinals next week; recall that I do a long engine team WOD every Saturday”

I wanted to create a workout tracker that can handle serious training demands. I ended up with a real-time AI companion at your fingertips that can guide you like a real coach.

There’s one more piece I haven’t talked about yet. All of this data — your training history, your recovery trends, your AI coaching conversations — paints a picture of who you are as an athlete. Your fitness level across workouts. Your strength relative to established standards. How much you’ve actually practiced each skill over your lifetime. I’ve been calling this ‘My Fitness’ — a gamified snapshot of your athletic identity that evolves as you train. Think of it as Whoop’s ‘real age’ but for functional fitness: one page that shows you exactly where you stand and what to work on next. You’ve seen it on the web and it’s coming to the native app soon.

“Omg, look at this beautiful design”, Cat said. “Finally, someone made a beautiful period tracking app”.

I was looking at a stylish “28” logo and UI with a color palette that could easily win Apple’s app of the year.

It had not just tracking, but symptom logging, workouts with videos, diet suggestions and recipes; a whole ecosystem, aka cycle-syncing.

Her and I have been half-jokingly talking about creating a cycle app that doesn’t suck. The app we were looking at surely looked like one: it had a ton of functionality including much-desired partner integration, the UX made sense, and a gorgeous design tied it all together.

“Wow, someone finally vibe-coded what we’ve envisioned”, I thought.

I fired up deep research. Few minutes later it printed out something that I didn’t quite expect: “Started in 2019... two founders who were working in brand...” Before my surprise could wear off, it was replaced with a familiar thought: “Overnight success doesn’t exist.” I’ve seen this so many times yet I’d still fall victim to thinking that someone just came up with something brilliant over a hackathon weekend.

“Dude, you gotta teach me how to be so good at ring muscle-ups”, Dee told me as I finished my 15th rep on Open 26.2. Dee was a level 2 CrossFit coach at the box I was visiting in Vietnam. The 2nd workout this year ended with a brutal set of 20 ring muscle-ups after about 12min of other shoulder-intense movements and high-skill gymnastics. Even doing 1 muscle-up was a worthy achievement and put you straight into 75th percentile.

My score put me in 90th. Out of 20 RX participants in our gym, only 2 others did better. It seemed I immediately gained respect among everyone as half of the box watched me perform perfect reps as the timer ticked down.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

The next day I came across a ton of Reddit posts on Open performance frustrations. Folks were devastated with their scores.

This reminded me of my own feelings few years ago: I was doing powerlifting and bodybuilding -style training for many years. I thought I was strong and jacked and fit and all the things. Then I joined a CrossFit class and finished one of the last. In the years that followed I’d routinely perform way below what “good” meant by CrossFit standards. I was humbled again and again.

Don’t just take my word for it, here are my Open results to prove it:

What people don’t see leading up to those flawless ring muscle-ups are all the times someone practices them consistently, for weeks, months, or years. What they don’t see is the slow growth from bottom 5% to to top 17% over 6 years. No overnight success here either.

When I started building PRzilla last year, I surveyed over a dozen of CrossFit boxes and their coaches. I wanted to see what they like/dislike about their existing platforms like Wodify or SugarWOD. One theme was common—some gyms recently switched to PushPress, a “new kid“ on the block—and they were all quite happy with it. It provided a no-nonsense, all-in-one class management solution as well as social features for members. I remember checking it out couple years ago when I dropped into a box in Bangkok: clean, simple UI, modern features like a social feed. It looked a lot like what I would build.

Fast forward few months and I’m chatting with . “PushPress is 12 years old, yet I still come across affiliates who have never heard of it”, Dan tells me. I want to say I’m in disbelief but, of course, I’m not. I know better. “We were at 0 profit first few years and I almost quit. And now it feels like we’re only getting started.“

As I’m nearing a year of working on my vision, it’s easy to want to be so much further. More daily visitors, more registrations, more app installs, more features. Every day you see incredible products that were bootstrapped in “just few months”, or vibe-coded in a weekend while talking to Claude on the phone from the subway.

Every day I see 2-3 apps that claim to do the same thing as what I’m building. Shouldn’t I be further? Shouldn’t I be better?

And so I remind myself: there’s no overnight success. The grind continues. Step by step. 1% better every day.

TLDR: xenom calculator is here

The other day I came across XENOM — a newly-founded global CrossFit competition with a known and fixed set of workouts. Think CrossFit Games but standardized to a consistent two-day event similar to HYROX. You get a score on each of the 10 workouts and your final result is the sum total.

This got me thinking: can we determine athlete’s performance on this kind of event?

In HYROX world, people ask the same question: what’s my estimated HYROX time if I’ve never run the race? Why do we care? Well, if we know user’s estimated time we know their starting level, their weakest splits, best division to compete under, and—most importantly—what improvement is realistic and what exact program to follow to get better for the upcoming race. Just like to run a marathon your training would look very different if you’re at 8min/mile with 10 miles weekly average or 5min/mile with 50miles/week.

I’ve been working on CrossFit benchmarking with AI for a long time. But the XENOM problem is different: instead of figuring out time/reps on a given workout, race estimate is asking us to map athlete performance onto a score on a different workout.

So how do HYROX calculators do this? Because HYROX is a sequence of single movements, its math is a lot simpler. Take your 1km running time × 8, add time of each of the stations (rowing, wall balls, etc.), and add fatigue multipliers for each.

XENOM estimate is asking: how well would I do on 4 attempts of max snatch in 9min, on a wall walk + rope climb ladder, on a 12min WOD ending with max rep muscle-ups, on a 3km run into 2k ski, and so on.

Obviously the best way — outside of doing an entire mock race — would be to just attempt each of those workouts and plug in your results1. But in the age of modern AI and having access to your training data (like we do in PRzilla), we should certainly be able to figure this out without actual attempts.

1RM snatch (XENOM 001) is easy: just plug your best max lift. Recent snatch matters more than your all-time-best from 5 years ago.

What about more complex workouts?

Putting my coach hat on, to determine the score on XENOM 002 which is an ascending ladder of 2 wall walks + 1 rope climb, 4 wall walks + 2 rope climbs, etc. for 8min, I could start by checking user’s recent wall walk performance. Ok, they’ve done 5x5 wall walks a month ago and 3 rope climbs a couple months ago. They’re capable of getting to 2+1, 4+2 and can likely get to 6+3 as well and perhaps even 8+4. I’m able to predict this because 5x5 tells me that 5 is not their true max (1RM) but likely 7-8 is.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Here’s the thing: max consecutive reps are not a perfect predictor to what an athlete is able to do. Some aren’t able to piece 10 wall walks in a row but they can bang out 30-45 via 10-15 sets of 3 with short rest. That’s a first wrench in our calculations. I might see that you’ve only ever done 3 consecutive wall walks, but I don’t know what your muscular endurance on them is. A classic example is Cindy — if you did 20 rounds RX I can say with high confidence that your muscular endurance is quite strong; you’re able to continuously perform 5/10/15 of pull-ups/push-ups/squats for 20 minutes. Same thing with wall walks: being able to do 10 in a row is an indicator of your continuous-set endurance and proficiency in the movement... but an even better predictor could be looking at a WOD that includes plenty of wall walks like Open 21.1, 22.1, 23.3, etc.

As a coach, I can also look at your handstand push-up performance, either as a max-rep number or in a WOD that includes them (e.g. Diane). How about handstand walk. You can do 50ft? Your handstand endurance is generally “strong” and so your wall walk performance should be at least L5. But here’s a curve ball: can you extrapolate handstand walk performance from a handstand push-up? The stabilizing muscle stamina has some overlap but you can certainly get good at handstands without ever being able to do a handstand push-up; the latter requires strength and stability throughout the entire range of motion.

Even if we have your training data, CrossFit-style workouts make our calculation hard because you log a score, you don’t log how you performed the movements.

In a WOD that calls for 3 rounds of 20 C2B, did you do them as 5 sets of 4 because the mastery isn’t quite there (e.g. ~L3 perf) or did you bang out sets of 20 (e.g. ~L8 perf) because you own them? We can somewhat extrapolate it from the total time: an athlete that’s able to do easy 20 is likely to finish workout faster than the one breaking in 4 sets, but it’s not a direct indicator of your movement performance.

Thankfully, when logging untimed practice (standalone movements) in PRzilla, you log them as sets and reps. This makes it easier for us to determine your max reps without a standalone test. Just like logging 3x10 bench press @185lb tells us your estimated 1RM.

There is a good opportunity for disruption here: specify sets/reps when logging WODs for that ultimate analytics.

Another curveball: let’s say you log 10 rounds on a WOD that’s a 12min AMRAP of 5 deadlifts and 5 wall walks. This is a good indicator of your capacity but did you do 10 rounds because of deadlift strength or wall walk strength, and which percentage of each contributed to the final score? It could be 80%/20%—you’re a powerlifter with 500lb deadlift who’s never done handstand work; or 20%/80% — you’re a gymnast who’s never done deadlifts. We should be looking at a broader set of workouts: if you consistently do well on those with handstand movements, we can assume with more certainty that you’re good at them.

While we know that absolute strength corresponds to being able to do more work at a given weight, can we really be sure that a person with 255lb clean will do better on Grace than a person with 205lb clean? I’ve seen guys in the gym who never go above 185lb but they can cycle 135lb forever, and do it fast. On the other hand, my clean is closer to 225 but my HR is through the roof after 15 singles with 135lb. In other words, you don’t need to push the ceiling in order to get good at sub maximal weight endurance2.

This is why looking at a user’s DT and Grace scores is as important as looking at their Clean 1RM; it shows their performance in barbell cycling, submaximal strength and endurance rather than their absolute max.

Why is Grace such a great benchmark? Because it asks you to perform X reps in one movement in shortest time. This is similar to a famous 30 muscle-ups for time. If we invert this into max reps in X time, you have tests like Handstand Push-ups: Max reps in 2 min. Going back to wall walks, if we know max reps athlete can complete in 2 min, that’d be one of the best proxy benchmarks for events like 002.

An interesting proxy that exists in HYROX world is their Physical Fitness Test: run 1000m, do 50 burpee broad jumps, 100 stationary lunges, 1000m row, 30 hand-release push-ups, and 100 wall balls.

15–25 min is PRO, 25–35 minutes is Open, 30–40 min is Doubles, 35–45 min is Relay.

Notice anything? PFT is similar to Grace or 30RMU’s: it’s a chipper of all of the HYROX movements where you perform X reps on each as quickly as you can. This is a direct indicator of strength/endurance/capacity on each of them, making your level approximation quite accurate.

So should XENOM have its own PFT? HYROX’s works because it IS the race, just miniaturized. Same movements, scaled volume, done. Clean and obvious.

XENOM can’t do that. Ten events, nineteen movements across three completely different fitness domains — a miniature version is just a giant chipper. Technically, it would be something like King Kong or Fight Gone Bad. Practically, you’d be testing too many things in a way that’s not really representative of individual events: stimulus of a ladder is very different than a stimulus of max-weight attempts or a long endurance grinder.

We could come up with something like this: accessible to a majority of CrossFitters (no muscle-ups, no heavy snatch, no max-cal bike) but it would only be a faint projection of your overall performance.

For time:

1,000m Run
15 Thrusters (60/42kg)
15 Toes-to-Bar
1,000m Echo Ski
10 Cleans (80/55kg)
10 Handstand Push-Ups

Elite: <15 min, RX: 15–22 min, Compete: 22–32 min

In the meantime, I whipped up XENOM calculator in PRzilla.

Calculator is smart. It uses WOD scores that serve as best proxies for an event — DT, Fran, Amanda, 5k run, etc. If there are no benchmarks it analyzes your training data for relevant signals. It decays benchmarks at different rates based on scientific research: strength reduces slowly, endurance diminishes fast, and acquired skills mostly persist.

Here it’s showing that I’ll probably land right at that fabulous 50th percentile :) You can override each score if you performed that specific workout or feel like our projection is incorrect.

If you’re logged in, it uses your existing WOD scores under the hood. If you don’t have account, just plug your WOD scores manually and it’ll use the exact same smart calculations. The more scores you give the more accurate final prediction is.

I’ll be refining this calculator as we learn more about benchmarks. I’d like to add a division estimate to help folks decide what track to compete under.

Next version: I’d like to try feeding this through LLM asking it to reason as a coach. Coach doesn’t just run formulas — they read between the lines. They notice you've been logging at 70% for two months, that your shoulder-heavy movements have quietly disappeared from your logs, that your snatch PR is from 2022 but you've been crushing Isabel lately. That kind of contextual reasoning is hard to encode in rules.

Drop me a note if you have thoughts on this or just found it useful.

Many HYROX calculators do a very simple math where they ask you to put time of each station.

For more on strength and submaximal strength, see coach Shawn’s recent excellent writeup on this.

A big shift last year has been towards skill learning and widening movement repertoire. Now that I’m 40, I can’t quite recover as fast after going hard on WODs. I’ve also grown weary of constantly chasing higher strength numbers: squat, deadlift, clean. Most of those have plateaued years ago and now need a dedicated multi-month cycle to make meaningful progress1. And so I found that learning new movements and practicing existing ones is one way to improve fitness while giving body a break. It’s also fun and rewarding as you see your quantity and quality go up.

I’m now increasingly seeing my time in the gym as existing in one of 3 buckets.

I’m testing if I perform a for-score workout and push myself to my max or near-max. This is the most fun one, of course; it’s rewarding to end up in top 3 on a leaderboard. But it also comes with the highest cost. For muscular stamina workouts like Cindy or Eva, you’re very sore for the next few days due to the density of work (volume/time). For near-max lifts, your CNS is fried and joints take a beating.

I’m training if I perform a prescribed body of work at a specific, usually RPE5-8 intensity. Training could be timed but often isn’t. 40min EMOM with ~30sec rest each min is a good example. So is doing 10 sets of 3 snatches at ~75% with 1-2min rest. These are usually less fun, but they need to constitute most of the time.

I’m recovering if I consciously limit my work to a very low intensity and/or perform low-impact movements. This is where Zone 2 training comes in, such as spending 45-90min on a bike or a rower at ~130bpm (for my age). I managed to make these type of days less boring by throwing skill learning into the mix; I would do some handstand walks every 250-500m on a rower, or do 1-2 ring muscle-ups. Anything that’s ~20% of your capacity and ideally bodyweight-only.

Ideal week then looks like 1-2 days of testing, 3-4 days of training, and 1-2 days of recovery. This looks easy on paper but tends to be a very delicate balancing act and is quite difficult to get perfectly right.

The moment I push my testing a bit too much, I get injured. If I don’t include recovery, I stop progressing or quickly overtrain. Another takeaway: just because I can ace that workout and beat my (or other) scores, doesn’t mean I should.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Last year I’ve probably spent 30% on each of the following:

• High level gymnastic skills — Ring Muscle Ups and Handstand Walk

• Zone 2 training

• Longer/endurance -heavy WODs

RMU’s and HSW were the last two pieces of the CrossFit puzzle. It was what I needed to be able to RX 99% of WODs in a typical class.

I went from being able to do 2-3 muscle-ups to hitting 10 in a row, being able to do 4-5 on any given day, and under fatigue during WODs. I even accidentally progressed to 2-3 strict ones! Handstand walks were also something I struggled to not fall just a few feet in, and now I can almost always go 50ft, confidently. The power of practice truly can not be underestimated: all I did was show up and practice every week, sometimes once, sometimes twice. In that regard, the year has been a success.

The other part was focusing on improving my endurance: weekly zone 2 training and grinding out >30min WODs. The year ended on a good note as I finally completed Eva RX at 44:04. Earlier in 2025, I’ve done it with 24kg bell in 40min. Even if I didn’t improve in my fitness, it feels like my mental fortitude has gotten better with these grueling long workouts2; I’ve learned to suffer for longer and under more fatigue.

It’s hard to tell if my Zone 2 training is having an actual impact on VO2 max but it certainly feels like I can sustain on med/long -duration WODs better. Running also feels the best it ever has.

I walked into our box yesterday and saw a “Q1 2026 Goals” board. Someone wanted to hit a 210kg squat. Below were: “3 muscle-ups”, “15ft handstand walk” and a “kipping pull-up”. Last year I wrote “10 RMU” and “50ft handstand walk”. Few years ago I would have written 365lb (2xBW) squat or 185lb (1xBW) shoulder press; now all conquered.

I starred at the board, unsure about my next goal.

I’ve hit most of the things written on it and much more. It wasn’t about feeling superior; I was reminded of how far I’ve come and how many are still on their way to reach goals that I’ve conquered long ago.

Do I really need to work towards hitting 15 consecutive RMUs? Or 2x50ft handstand walks, which would be a natural progression. I’ve hit the point of diminishing returns. Being able to walk over obstacles is a semifinals territory and isn’t something I’d ever need in a regular class. To play devil’s advocate, hitting strict ring muscle-up isn’t something you’d need in a class either! Yet I wanted to be able to do it as a pure form of incredible upper body strength and control.

As I’m nearing 15 years into my fitness journey, all of this feels like an “advanced lifter” curse. When you’ve reached 90% performance on most things, the remaining 10% start to take a lot more time and the progress slows down to a crawl.

One option is to keep pushing towards Level 3 and 4 on this CrossFit standards sheet I’ve been using for the last few years. But while most of the Level 3 are achievable with a few month of dedicated practice, I find that many of the Level 4 would take years.

Genetics aside, the reason I hit 2.5xBW deadlift was because I powerlifted for a few years. Similarly, the only way to run 5min mile is to be… a runner. To go from my current 6:30 best (and now, likely, closer to 7:15), I’d need to start running 2-3 times a week, build capacity over a year, slowly improve speed, slowly adopt my body to volume, train the body to go that fast for that “long”. You don’t just run 5 min mile. You only do that if you’re a runner. And so the question then becomes: do I want to put all my eggs in one basket? How would running 2-3 times a week affect my strength, my muscular endurance, my gymnastic proficiency, and other aspects of fitness?

Biological age matters more than chronological one3, but it’s hard not to think about turning 40. Relatively, my body has less work capacity than an average 30 yo athlete of the same “level”. CrossFit Open also acknowledges it, putting me in a 40-45 bracket. Should the goal this year (and going forward) then just be to maintain existing fitness and avoid major regressions?

I’m already doing that intuitively. I’d make sure to squat heavy at least every other week: I’m way below my all-time-best 365lb but even hitting 300lb right now feels like a good baseline and puts me in the top 5 in the class. I might not be able to do 100 dubs like I did at one point, but comfortably banging out 30-50 goes a long way in any workout.

I realized that reaching a certain goal serves as an overextension that builds a ceiling. Hitting 10 RMU’s last year was a nice goal but the ultimate “friend we’ve made along the way” was the fact that 3-5 started feeling like nothing. To go back to running analogy, training for a 5 min mile would create a baseline where running 7min mile feels like a child’s play. I’d just need to keep that baseline by running once or twice weekly.

I ended up putting “Open 2026, do better than last year” on the whiteboard. It’s a fun short-term goal that would certainly push my engine further, but I’m very much aware that Open can be easily gamed and is just a subset of overall fitness and what CrossFit wants you to achieve. Next two months I’ll be focusing on frequent practice of top 5 open movements that are skill-sensitive: dubs, thrusters, muscle-ups, c2b, and rowing. I will also double-down on nausea-like WODs that Open is famous for.

I think if I were to pick one goal this year, it would be to continue building the engine (= endurance). I’ve built strength over many years and I’ve always been good at gymnastics (thanks to yoga since early age). Endurance is still lacking so I need to keep hammering >20 min WODs, more running, more hero ones.

Earlier last year I did Vipassana and it was pretty life-changing. I’ve noticed a lot more calm during intense workouts where the feeling of “dying” (zone 4-5) still feels awful but at least it doesn’t provoke as much of a panic state as it would before. I’m able to sustain in red for longer.

One of the intentions for this year is to maintain my meditation practice as it helps in both, “life” and when on the training floor.

Although I did finally hit 185lb (1xBW) snatch this year!

Another one I’m particularly proud of was 8 rounds of: 5 RMU, 7 DL (225/155), 5 Toes Through Rings, 7 Double KB Snatches (53/35), 5 HSP which I RX’d in 33:13

Whoop says I’m 4 years younger and Function Health says I’m only 30yo

As I was working on PRzilla — an app to track WOD scores, fitness skills and standards, I quickly realized that I needed a way to log new sessions without constantly doing a clunky, manual SugarWOD / Strong app export/import.

There was really no way around it — I had to build a “workout tracker” 🙈

The thing about most CrossFit tracking apps — SugarWOD, Wodify, PushPress — is that they’re focused around gym programming. This makes sense since community is at a core of CrossFit’s methodology. You sign up to the gym, attend classes, and never have to worry about what to work on, the workouts just appear in your app.

On the end of the spectrum there are apps like Strong and Hevy which are geared towards traditional strength training. They’re your classic workout trackers for individuals walking into the gym, hitting 3x10 bench press and 3x15 tricep extensions because they saw it in a Men’s Health magazine.

I come from the “left”, which I did for many years, using Strong app since 2018 (and Fitocracy before that… RIP). Then I started doing more CrossFit-style workouts and I’m now signed up to a box that uses SugarWOD. Couple times a week I join the classes if they look fun or include something I want to work on or… I simply want to workout with friends.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

As such, I’m often somewhere in the middle: I want to be able to come to the gym on Monday and log my own workout of, say, 10x3 ring muscle-up practice and 5k row that I do in between RMU’s, followed by a WOD like 15min AMRAP of 1,2,3,… shuttle sprints and wall walks. And I want to come to the class on Tue and log whatever SWOD, WOD, and other things they’re doing.

Now here’s the crazy part — there seems to be a bizarrely large separation between these two camps. PushPress, Wodify, and SugarWOD all make it really hard to log individual movements. You’re either doing Fran (and logging your score) or you’re shit out of luck. Wodify and PushPress are the most restrictive — you can’t even use them without a gym subscription. I get it, their business model revolves around gyms paying fees that depend on number of users. SugarWOD is a bit more relaxed in that you can switch between gyms (as long as you know their invite code).

All 3 support custom logging but it’s incredibly limiting and is always tucked away behind multiple menus. This use-case isn’t what they’re optimizing for.

Here’s Wodify help docs showing 5 step (!) process on how to log 2 sets of muscle-ups for the day.

PushPress isn’t too far off:

In SugarWOD, you need to click into “More” → “Logbook” → “Log a workout”, then chose from a limited list of weightlifting, endurance, or gymnastics. When I say limited, it means you can’t even log 10x3 of RMU (it mostly has “max reps” wods).

One exception to the rule is BTWB app that’s often considered the golden standard for CrossFit tracking. That’s why I put it closer to the middle in my diagram. It allows you to log individual movements as well as any WOD— custom or canonical—and it doesn’t lock you into a specific affiliate.

But while its “New workout” starts off nicely with “Single Movement” (this is your classic Hevy/Strong logging) vs. “Multiple Movements” (this is your classic WOD), it quickly transforms into a clunky UI that resembles a cockpit of an airplane, drowning you in a myriad of options that are frustrating to comb through:

I got to the gym and I have no plan. I decide to start with some light snatches… just give me a quick way to start adding the sets. With Strong, it takes exactly 2 seconds and 3 clicks: “Start an Empty Workout” → “Add Exercises” → Type “Sna” → “Add” and you’re all set.

If I want to then make those sets into a WOD, perhaps there should just be a way to do that. And if instead of an individual movement I want to throw in a quick WOD next, after my untimed practice, perhaps there should just be a quick option for that; and this is where I’d chose from AMRAP, For Time, For Load, EMOM, etc.

BTWB is close but is not quite that perfect best-of-both-worlds app.

Wodify docs on how to add custom workouts list some great examples for why you’d even want to log them:

I’ll add few more:

• What if I’m injured and can’t do today’s programming?

• What if I simply don’t like today’s programming?

• What if I feel kinda off today and just want to get on a bike and do some light cardio?

• What if I’m trying to get better at something so I’m substituting today’s programming or supplementing it?

• What if I completely switch to another affiliate programming because it serves me better?

• What if I stopped coming to an affiliate and switched to a home gym? Or moved to another city or country?

BTWB is the best of all evils here, if you can get past their awkward UX. But all the other apps are simply not great options for long-term CrossFit tracking. Most importantly, it’s important to be in control of your data: WOD performance and all the individual movement history.

An ideal app should allow to:

1. Log individual movements aka Strong/Hevy; plain sets-and-reps/time/distance, quick and easy.

2. Log a score for a canonical WOD (Fran, Murph, Open 21.4, etc.)

3. Log a score for a custom WOD (whatever you came up with — AMRAP, For Time, For Load, EMOM, etc.)

4. Subscribe to any box, see their programming, log their WOD’s as part of your own data! This is really just 2 or 3 but the WODs (canonical or custom) are provided by the gym.

Bonus points if you can specify exactly how you scaled a WOD so that the data can be analyzed programmatically. Adding “25# db” in notes doesn’t count since we can’t really measure your progress and your effort!

After couple months, I’m now pretty satisfied with the workout tracker in PRzilla. I’ve been using it exclusively for a few weeks and have gotten to the point where I no longer need to use any other app.

I’ve learned that building “simple” workout tracker isn’t as trivial as I thought 😅 Despite engineering for 20 years I still grossly underestimate complexity of systems. The hardest task was figuring out active/draft vs. completed sessions as well as various constraints: there could only be 1 draft at a time, there could only be 1 workout session per day, etc.

Then there’re hidden things like ranking movements in a search list so you get most relevant show up at the top. It’s what separates good UX from the great one. I also use supersets almost daily so I came up with a UX to link cards together via drag and drop — something I haven’t seen in other apps.

The best part about building your own app is that you control an entire experience and can “easily” implement any features you need. E.g. one idea I have is to show inline insights about workout sessions after you’ve done them: total volume lifted or distance ran, relative effort (based on your 1RM), milestones (“you’ve just reached 1000 miles cycled this year” or “this is your 2nd best consecutive ring muscle-up set”), etc. This could be shown every time you log a workout.

I can now also easily create goals and display them as progress bars. Perhaps you’re working towards a 300lb bench press: we can show your progress easily. Or maybe you’re working on reaching 100 muscle-up sessions this year: we can show that as well.

The things I haven’t added that exist in Strong/Hevy that I might consider in the future:

• Movement reordering (easy to add but it needs to work well with supersets)

• Rest time between sets

• Template-like logging; i.e. adding set in UI is not the same as completing it

Stay tuned on the future developments and if drop me a note if you’re using it and have feedback (or would like to have all of your SugarWOD data imported).

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

My fitness journey—as is the case with many teenagers—began with bodybuilding. I wanted to look good. Soon after, I found StrongLifts 5x5 and got into powerlifting. It became all about numbers: getting bigger bench, bigger squat, bigger press. Yet, I’ve always been drawn to the notion of Total Athleticism, as coined by Max Shank in one of the articles on T-Nation that I used to read religiously around 20151. Having a 300 bench was cool but I didn’t want to be one of those powerlifters who had massive numbers yet couldn’t run up the stairs. I wanted to also be good at running, calisthenics, kettlebells. Eventually this brought me to “functional fitness” and, of course, CrossFit, which popularized it circa 2000.

In my fitness circles, CrossFit was still criticized for its reckless high-skill olympic movements performed at high intensity. Blame the epic fail videos of someone doing something stupid and the ignorance around the methodology. While I was on the offense about doing actual CrossFit, I loved the “variable movements” concept. I found a couple of “crossfit athlete standards” posters online and made this spreadsheet to track my progress across multiple domains. It immediately exposed all my gaps: I could squat 2x bodyweight but my snatch was at a measly 100lb and all the speed and work capacity tests were barely at level 2:

If I wanted to be an all-around developed athlete, these were the things I had to work on. The standards also served as an “objective” benchmark. To consider yourself “advanced” here’s how many pull-ups you had to be able to do; and this is how fast your 1 mile run would have to be. It gave me a concrete goal to work towards. These spreadsheets became my north star for the following few years. For a challenge junky like me, they were a perfect long-term obsession.

The spreadsheet overload was real. This wasn’t the first one I used. As far back as 2011, I found exrx.net Strength Standards and created this view to understand where I stand strength-wise and what to work on:

In the last couple years I started tracking my frequency and total-lifetime-session-count of certain movements I wanted to be better at — a concept I wrote about before:

Finally, I tracked my proficiency levels on various CrossFit -specific movements as a way to advance my skill and become fluent in them during WOD’s.

The year was 2025.
I was a software engineer.
And I would still manually update a spreadsheet with the number of times I’ve performed a certain movement that I deemed as “needed practice“.

This was embarrassing.

When I embarked on building PRzilla, I realized that perhaps this was the time to ditch manual spreadsheet tracking. I could now build an app that would have all of this backed in:

• show your fitness level across multiple movement patterns/domains (strength, endurance, gymnastics, work capacity, etc.)

• show your raw strength benchmarks (squat, deadlift, snatch, push press, etc.)

• show your skill proficiency as a “lifetime sessions performed”

  • if you’ve done ring muscle-ups only 20 times in your life, you’re unlikely to be better at them vs. someone who has done them 120 times

• show your skill ownership as a “max consecutive reps able to do“

  • being able to do 50 consecutive kipping pull-ups means you own them; this movement is unlikely to be your limiting factor in any WOD that has them

I use Whoop and I absolutely love how it’s able to distill complex/boring HRV/RHR metrics into simple, quantified scores like recovery and strain. Whoop and Apple Fitness—that’s just as big on quantifiable fitness—were a big motivation for this app.

On the other hand there are apps like BTWB which is one of the most extensive Crossfit-style workout tracking tools, but I found its UI unintuitive and UX clunky:

And so I turned all my spreadsheets into a simple snapshot consisting of 5 views: your level, balance, strength, ownership, and practice. These could be easily extended in the future with any other modules: time domain distribution, specific goals tracking like work capacity or endurance. Or even sport-specific ones like Hyrox.

In order to turn all my workout data into beautiful charts, I needed to… have that data in the first place. One issue was that it was split between:

• SugarWOD — scores of WODs prescribed by my box that I did

• wodwell.com — common WODs that I did on my own

• Strong app — traditional strength training workouts that are not WODs (aka sets and reps)

Importing wodwell scores was easy since it was just a map of common wod (fran, murph, etc.) to a score. Strong app export would be a lot more involved since I would have to implement sets and reps tracking (as well as workout sessions, potential rest values, and so on). So I decided to focus on SugarWOD import. And this is where the fun began.

Good news: SugarWOD allows easy export of all of you workout history.
Bad news: SugarWOD data is very… unstructured.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Here’s an example of CSV:

09/14/2022,WOD,24 Minute AMRAP:Row 240m12 Lateral Burpees over Back of Rower48 Double Unders24 Alternating Front Foot Elevated Reverse Lunges (53/35)*,3.073,3+73,Rounds + Reps,,"[{""rnds"":3,""reps"":73}]",,SCALED,

As you can see, we have an arbitrary, potentially non-descriptive wod title like “WOD” plus gobbled up, plain-text description like “24 Minute AMRAP:Row 240m12 Lateral Burpees over Back of Rower48 Double Unders24 Alternating Front Foot Elevated Reverse Lunges (53/35)*” that’s missing basic formatting / newlines.

As humans, we’re able to quickly parse this into:

24 Minute AMRAP:
Row 240m
12 Lateral Burpees over Back of Rower
48 Double Unders
24 Alternating Front Foot Elevated Reverse Lunges (53/35)*

Thank god we live in the age of LLM’s which are capable of reasoning through a messy jammed up text like this just like we—humans—do.

Another peculiarity was load-based entries. In SugarWOD you can program them in a workout and specify sets and reps, e.g. 5 sets of 3 snatches. Users can then log a value for each of the 5 sets. In order to present your performance on the leaderboard, SugarWOD allows coaches to specify how to score those sets — max value (who got the highest weight)? lowest value (who got the fastest row time)? sum of all values (who did the most work overall)? and so on. The export doesn’t expose this scoring criteria so parser needs to infer it based on the sets data. In the example below, we can see that the scoring was using SUM of 12 sets and so 695 is not the weight user did as a 1RM squat snatch; the real squat snatch values are in the sets field:

05/31/2024,WOD,12 ROUNDS:30 Second CAP:3 Toes to Bar2 Lateral Barbell Burpees1 Squat Snatch*REST 1 Minute.*Increase weight as able.,695,695,Load,,"[{""success"":true,""load"":55},{""success"":true,""load"":55},{""success"":true,""load"":55},{""success"":true,""load"":55},{""success"":true,""load"":55},{""success"":true,""load"":55},{""success"":true,""load"":60},{""success"":true,""load"":60},{""success"":true,""load"":60},{""success"":true,""load"":60},{""success"":true,""load"":60},{""success"":true,""load"":65}]",,RX,

In this case it’s “obvious” that 695 wasn’t a 1RM snatch (current world record is 496lbs) but some cases are much less obvious so the parser needs to be very careful there.

And so after many weeks of experimenting, refining, rewriting, and adjusting based on real data (thanks to amazing volunteers in my gym), I now have a pretty smart and capable pipeline that turns unstructured SugarWOD data into a structured PRzilla data:

One of the biggest findings—and things that slowed me down—was realizing that a giant, monolithic LLM prompt we used to generate giant JSON with a dozen of different fields (gpp, modality, difficulty, benchmarks, classification, etc.) was taking way too much time, was way too expensive, and often produced errors as it tried to do too many things at once.

I then switched to a series of small, targeted LLM parsers/prompts—as seen on the diagram above—for each of the metrics and ran them in parallel. The results were astonishing: faster and cheaper execution and much more accurate results. This also gave me flexibility to run only specific parsers in specific cases; e.g. when parsing your historic data we want to extract movements (to feed it into our proficiency metrics), performance levels (to understand how your fitness level progressed), time domain (to see a time domain breakdown), and so on. We don’t care about coaching/scaling/stimulus module since the workouts are in the past and users don’t need to know that! However, those modules are important for new workouts, when using analyze or generate.

Finally, it allowed me to run these parsers in parallel which meant that a WOD analysis was now taking time_of_slowest_module (usually ~12-15sec) rather than sequential SUM(module1, module2, …) that would usually take up to 40 sec!

The end result is incredible. We’re able to turn a textual mess like this, into an actual workout and your actual performance. Here we see that 14min AMRAP was properly parsed into movements like “Wall Walk” and “Front Squat”; that it’s an endurance and stamina -heavy workout (yep!), that it’s classified as “Very Hard” and its modality are equally “Gymnastics” and “Weighlifting”. Moreover, AI determined that user’s score of 80 falls right around L3 (which we would likely adjust to L3.5 or L4 due to workout’s difficulty):

Parsing “Wall Walk” as a movement is what allows us to count that towards your practice score! Notice that we now know that you’ve done “Wall Walk” 32 times in your life with the recent one being 6 months ago.

Now that we have this structured data, the possibilities—all of a sudden—are endless. We can easily, and more importantly, automatically show your strength levels: powerlifting, weightlifting, crossfit total:

We can derive how good you are at Endurance, Stamina, Power and other GPP components based on your scores on WODs that are high in those:

Because we’ve determined time domain of all the wods you’ve ever done, we can show if you’re leaning towards shorter or longer ones. Yes, parsing 1300 entries is expensive but at least we can marvel at the end result 😅. Here is coach Mike’s real data dating back to 2018. You can see that early years prioritize short WODs (<12min) whereas last couple years the focus has shifted towards longer, HYROX-style ones:

And, of course, we have ability to see all the WODs for any given movement (why it’s so important to parse those for all the custom WODs and create proper associations). Here you can see that Mike has done over 232 lifetime front squat sessions over 8 years, 157 as dedicated lifts and 75 as part of WODs:

In the interest of brevity, I’ll stop right here. There are other things powering this pipeline which I’m still refining and perhaps can talk about later: male vs. female benchmarks, age-based adjustment of strength and fitness metrics, smart movement aggregation for practice skill screen, logging import errors like movements that don’t match in our DB, or a smart system of retrying LLM when parsers fail.

Now that I’ve gotten here, I can’t help but wonder: what’s next? and what’s the end goal? I can now replace spreadsheets with this app but it doesn’t solve all of my use cases. The dream would be to have an app that can track all of my workouts. This means:

• It needs to be a native (mobile) app

  • Web apps are great but when in the gym and on the go—let’s be honest—we all prefer native apps.

• Replace SugarWOD completely?

  • PRzilla is able to do this by parsing previous data but what about future one?

  • I would need to either:

    • Implement SugarWOD API integration that’s tied to a box directly.

    • Implement some sort of image recognition of a WOD (snap a TV in your box) that can then be logged directly into our system.

• Allow custom sets and reps logging

  • This is a big one… and it would allow me to switch completely away from Strong app.

  • But first… I’ll need to port Strong app data into our system (perhaps more on that in later posts!)

Alex Viada came out with Hybrid Athlete around the same time.

When I left LinkedIn, I itched to build something in the space dear to my heart — fitness and CrossFit specifically. I also wanted a challenge of building a full-stack app, something I’ve never done before. The app would solve my pain points but I wanted to release it out there for anyone to use. This meant database, auth, users, and production-level user experience. It would be the biggest project I’ve ever done. With the rise of AI-assisted coding, it was a perfect time.

I’ve been using SugarWOD to track scores for CrossFit workouts (WODs) prescribed by my gym. But SugarWOD was never designed to be a standalone tracker: it’s missing many WODs, those that are there don’t have any info, and it doesn’t have a way to discover new ones. So I supplemented it with wodwell.com to find more workouts and track their scores.

Wodwell has its own issues: full of ads, a clunky UI, and it's slow. More importantly, I wanted to be in control of my data and wodwell has no export. It also wasn’t great that my workout history and performance data was split between two platforms.

All of this sounded like a perfect opportunity to build just that: a full-stack app that has an incredibly easy and fast search through a 1000 most popular WODs. It would allow to log scores for any of them, to track your progression over time, and to favorite WODs for later. As I started coding with AI, I quickly realized I could go even further: we could get insights into WODs via AI analysis (time domain, difficulty, L1-10 benchmarks, etc.)

Before I set out on this journey I wanted to define few foundational tenets that were non-negotiable.

“Vibe coding” exploded as I was starting this project. My LinkedIn feed was full with “this is incredible” and “this will never work” posts. I came across Addy’s article on Cline and decided to build this app entirely with AI as a matter of principle. No manual coding. It would be a perfect experiment since an app was not just a trivial one-pager vibe-coded in a day.

Always a fun UI challenge and is certainly a must these days unless you provide a native app. In context of CrossFit, you often need to look things up or log your scores while in the gym. Every page needed to be responsive and every UI concept needed to be adopted to small and large screens.

Not a terribly complicated constrain and is largely solved by using the right foundational abstractions but it does add cognitive complexity, especially if you’re working with AI, as you need to ensure it complies and uses the right tokens.

Often overlooked aspect but it’s what separates a polished, predictable app from a clunky frustrating experience. URL’s are the source of truth. Important UI state change needs to be reflected in them. Now you have the power to reload it, bookmark it, share it, go back, and so on.

Next.js is known for SSR support out of the box; this means fast server-driven apps. This was a great opportunity for me to learn and experiment with these concepts.

Big lesson I took from introducing these in the beginning: each constrain is a liability, another dimension to your product surface. Be careful with creating too many from the start. Think iPhone and its lack of copy-paste first few years.

A feature alone is a single point or a line (1D).

• Add a "mobile-ready" constraint, and that line now exists on a 2D plane (feature x device). You have to test both states.

• Add "dark mode," and the plane becomes a 3D cube (feature x device x theme).

• Add "SSR-ready," and you're now in a 4D space.

I’ve now spent about 5 months working on this daily-ish. I learned a ton about AI assisted coding and wrote about most of it. The lessons never stop and I post them weekly on LinkedIn.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

What started as a simple way to see most common WODs, quickly turned into a powerful UI that allows to find just the right workout. With the power of AI, I’ve gone deep on classifying workouts to create helpful data that doesn’t exist anywhere else out there — difficulty, modality, training stimulus, time domain, and workout characteristics via tags.

When I ask AI to summarize the complexity 1 of the app now:

PRZilla is a large-scale production web application with 123,000 lines of
TypeScript code across 818 files, featuring 19 database tables, ~109 React
components in 304 TSX files, and 67 tRPC API procedures. The codebase
includes 1,532 test cases with 256 E2E tests across 40 test files ensuring
critical user journeys, 58 service modules handling complex business
logic including 6 AI-powered features, and manages 922 predefined
workouts with sophisticated scoring algorithms. This represents
approximately 2-3 years of full-time development effort, comparable in
complexity to a mid-sized SaaS product.

It’s incredible to see the kind of power you wield with AI. The breadth and depth of functionality certainly feels like it would have taken me 2-3 years. I haven’t written any of this code and honestly can’t imagine having to ever write code manually again.

Cutting wood by hand is slow. Using an electric saw freehand is fast, but it’s how you get a crooked cut. The real leverage comes when you bolt the saw in place at a precise angle, set the exact speed, and let it execute a perfect cut in a minute.

That is exactly how I build software now. I don’t write code manually. And I don't just hand a task to an AI. Instead, I architect the system, protect it with guardrails so it stays the course, and give it specific instructions so it knows exactly the path to follow.

My role has changed: I am the architect and the guardrail engineer.

Having spent a good amount of time not only developing new features but also refactoring, redesigning UI, and fixing bugs, I can tell with good confidence: your app will not fall apart. AI is capable of 95%. The remaining 5% are complex cases that usually reside at the edges of larger system integrations OR are just complex in nature. Those would be also complex for human, likely even more so.

For example, I’ve struggled to implement a well-working lazy loading of WOD cards on the main page because there was already a complex state management of various filters that had to all work in unison and support SSR; introducing lazy loading created X^Y^Z level of state management complexity and AI struggled to keep everything together without small bugs popping up here and there.

These are the fundamentally hard issues inherent to engineering. AI offers no magic wand for challenges like:

• The "dependency hell" of npm packages.

• The chaos of flaky end-to-end tests.

• Navigating features with no documentation.

AI also can’t make your app stable if the underlying structure is rotten: fragmented state, logic duplication, complex branches with subtle bugs. But it’s surprisingly good at finding those and fixing them in a heartbeat.

When I look at the app right now I feel like it would have taken me much less time to build the “final” version. Yet, the reality is that development works like this in non-trivial apps:

AI allows you to travel that curvy path much faster. Although you have to be careful because without proper guardrails you can start swinging too far left and right: you created too much code, too many experiments, pushed things to prod too fast, all leading to too much liability.

You develop a feature, you have 1 problem.
You decide to release it into production, now you have 10 problems.

Besides the app looking “good” and working “smooth”, the most important production-level aspect is making sure you don’t break things. In the last 10 years I’ve worked at big companies where, despite often being oncall, you always have dedicated SRE help. You also have a well-oiled infra machine to detect errors in prod and notify you.

Thankfully, for small full-stack apps like mine, platforms like PostHog & Sentry are incredible and provide all-in-one solutions for error monitoring (and more) with generous basic tiers.

I followed a pretty standard, tiered approach to release things safely:

• TypeScript must pass

• Linter must pass

• Unit tests must pass

• E2E tests must pass

• Test locally to ensure things work

• Always push to a branch in production (Vercel makes it easy). This is basically your staging environment since it’s hitting production DB.

• Manually test feature in prod branch, merge into main if it works well. An even safer option would be to introduce feature flags with gradual rollouts but I didn’t want that complexity just yet.

• Finally, watch out for spikes in errors following the rollout of a commit.

Big takeaway here was not to trust AI with E2E tests. I didn’t pay too much attention to all of the assertions at first, then quickly discovered that bugs weren’t being caught. Turns out quality of E2E assertions was subpar: tests relied only on visibility checks, many used vague assertions or hard‑coded values, and almost none validated data against the database. Tests were slow and flaky due to waitForTimeout calls and text-based or CSS-class selectors. I ended up adding lint rules (via eslint-plugin-playwright) to ensure AI doesn’t break this in the future.

I struggled with design at first but later found that it’s often a matter of the right prompt. For example, this was prompted to look like Apple Fitness / Whoop in 2025 with Sonnet 4 (which translates to clean, modern and minimal UI with oversized elements):

Compare to the old one:

To summarize, I think at least 80% of my time was spent on making things polished: figuring out UI/UX, refining UI/UX… endlessly, testing various permutations of an app, thinking through edge cases, ensuring it’s tested well, ensuring it’s feature-complete yet not over-engineered, documenting it well, deploying it correctly, and so on.

In the next post, I’ll dive deeper into some of the fitness-heavy concepts I’ve implemented in the app. We’ll talk more about that colorful “My Fitness” page and the complex LLM-powered pipeline that powers it!

Code metrics don't tell the whole story, but they do provide a rough idea of this app's scale. Recognizing that AI can introduce bloat, I carefully reviewed and streamlined all committed code. I estimate the result is a lean codebase with no more than 15-20% potential cruft.

First, the Buddhist terms which come from the chain of Dependent Origination (Paṭiccasamuppāda):

1. Vedanā (Sensation/Feeling-Tone): This is the raw, unprocessed, pre-cognitive feeling that arises from sensory contact. It's not a complex emotion, just a simple "pleasant," "unpleasant," or "neutral" tag. It's the first flicker of experience.

2. Taṇhā (Craving/Aversion): This is the immediate, gut-level reaction to vedanā. If the sensation is pleasant, the mind reflexively generates a "want more" signal (craving). If it's unpleasant, it generates a "get rid of it" signal (aversion). This is the pivot point.

3. Saṅkhāra (Mental Formations/Conditioned Reactions): This is where the reaction deepens and solidifies. It refers to our complex, deeply ingrained habitual patterns of thought, emotion, and behavior that are triggered by craving or aversion. You are correct that they are often negative patterns, as they are reactive and automatic rather than conscious and wise.

Goenka's teaching that "craving begets craving and leads to more and more misery" is a direct description of how this chain becomes a self-perpetuating loop.

Modern neuroscience reveals that this ancient psychological map corresponds directly to observable brain functions. The entire process can be understood as a powerful, habit-forming reinforcing feedback loop built on three key scientific concepts: Reward-Based Learning, Hebbian Plasticity, and Cortical Hijacking.

When a sensory input arrives, it is immediately tagged with an emotional valence by the primitive parts of the brain.

• The Mechanism: Information from your senses makes a rapid first stop at the amygdala, the brain's threat and salience detector. The amygdala instantly tags the sensation: "Pleasant! Pay attention!" or "Unpleasant! Danger!" This happens in milliseconds, long before your conscious, rational brain (the prefrontal cortex) has had time to process it fully. This initial, lightning-fast tag is the biological equivalent of vedanā.

This initial tag immediately triggers a chemical reaction that fuels craving or aversion.

• The Mechanism (Craving): If the sensation is tagged "pleasant" (e.g., the taste of sugar, the "like" on social media), the brain's reward system is activated. The ventral tegmental area (VTA) releases the neurotransmitter dopamine into the nucleus accumbens. This dopamine surge doesn't just feel good; it's a powerful signal that says, "That was important! Do it again!" This is the neural basis of taṇhā (craving).

• The Mechanism (Aversion): If the sensation is tagged "unpleasant" (e.g., a critical comment, physical pain), the amygdala triggers the release of stress hormones like cortisol and adrenaline. This creates a state of anxiety and vigilance, powerfully motivating you to escape the source of the feeling. This is the neural basis of taṇhā (aversion).

This is the core of the reinforcing loop. Every time you act on that dopamine-driven craving or that cortisol-driven aversion, you strengthen the neural pathway that produced it. This is a fundamental principle of neuroplasticity called Hebbian Learning, famously summarized as: "Neurons that fire together, wire together."

• The Mechanism: When you repeat a behavior in response to a cue (the vedanā), the connection between the sensory neurons, the emotional/reward centers, and the motor centers in the basal ganglia (the brain's habit center) becomes faster, stronger, and more automatic.

• The Reinforcing Loop in Action:

  1. Cue: You feel an unpleasant sensation of anxiety (vedanā).

  2. Craving: The brain craves relief from this feeling (taṇhā).

  3. Routine: You habitually pull out your phone and scroll social media. The novelty and intermittent rewards provide small dopamine hits, temporarily overriding the anxiety.

  4. Reward & Reinforcement: The temporary relief reinforces the entire neural circuit. The next time you feel anxious, the urge to grab your phone will be stronger and more immediate.

This is the scientific explanation for "craving begets craving." The very act of satisfying the craving digs the neurological groove of that saṅkhāra deeper, making it the brain's default response.

Over time, this reinforcing loop becomes so powerful that the prefrontal cortex (PFC)—the seat of rational thought, long-term planning, and self-control—gets "hijacked." The automatic, reactive pathways from the amygdala and basal ganglia become dominant.

• The Mechanism: The brain learns that the quickest way to soothe the discomfort of craving or aversion is to perform the habitual action. This creates a state of "dopamine deficit," where you need more and more of the stimulus just to feel normal. The "misery" Goenka speaks of is the scientific state of being caught in this loop: you are constantly agitated by craving/aversion and compelled to perform actions that provide only fleeting relief, all while your capacity for conscious choice and long-term wellbeing is diminished.

The practice of Vipassana is a direct intervention in this loop. By observing the raw vedanā (the pleasant or unpleasant sensation) with mindful equanimity, you refuse to feed the next link in the chain (taṇhā). You feel the itch but don't scratch it. In neurological terms, you are activating your prefrontal cortex to consciously inhibit the automatic, reactive firing of the amygdala and basal ganglia. By repeatedly doing this, you weaken the old, reactive neural pathways and, through the same principle of neuroplasticity, begin to build a new, wiser pathway of non-reaction.

I first heard about Vipassana few years ago. Also known as “10 day silent meditation”, it seemed like an extreme challenge for highly spiritual people detached from the regular joys (and miseries) of life. So… not for me.

Recently, a close friend has gone through one and highly recommended it. Blame middle-life crisis, but this time I decided to give it a try. “It could be a cool new challenge“, I thought. “How hard could 10 days of ‘no talking’ be”?

Little did I know how far off my understanding of Vipassana was and how impactful the experience would become.

So why do this? I’ve recently gone through a major life change — “quitting” my software engineer career of almost 20 years, not really wanting to go back to work for another company, searching for a deeper meaning other than making money and climbing the career ladder, and wanting to apply myself to something more purposeful and meaningful in life (= not involving helping billion-dollar company meet earnings expectations at all costs).

Perhaps related or in parallel to that, I started noticing myself living in the past a bit too much. Thinking about how nice things were “back then“, not really having anything to look forward to. I wasn’t enjoying life with the same intensity I once had. My brain was getting “solidified” and wired to existing connections; I needed to shake things up deep and from the bottom up. “Grumpy old man” had to go.

I got to the center in Delaware and learned that my next 10 days will be in the form of:

• Wake up at 4am

• Meditate for 6 full hours throughout the day (up to 16 if you so desire!)

• Spend the other 11 hours walking and thinking (no talking, no reading, no writing, no exercise)

• Sleep at 9pm

Fantastic… And so 10 days began.

During my time there I often wondered what surrounding neighbors think: a couple dozen people pacing back and forth through the park in the middle of the complex, sometimes standing still and starring into the horizon, sometimes talking to squirrels, sometimes sitting on the bench for so long — with *gasp* no phones in hand — that it would certainly be classified as “cuckoo“ in the “real world”.

Despite expecting some level of spirituality, I found almost none during the daily practice and discourses. Instead, it felt like we were part of an intensive training bootcamp but instead of physical exercises we were training our minds to be still, to not react, and to observe our deep body sensations. That’s it!

As a fitness aficionado, this really appealed to me. It was pure work and pure science. Meditation and sensory observation was something you could progress as you would with neuromuscular adaptations, “flexing” daily to become better, and it having tangible positive effect on your well being.

I realized that as much as we need physical movement, we need mental stillness.

Move your DNA postulated that we’re a function of our daily movements and daily behaviors in general. Since I read it, I started paying a lot more attention to how I move: using my left and right limbs equally, noticing tiny but deeply ingrained patterns that I could disrupt. This gave me a deeper awareness of my body. The more crude awareness came from years in the fitness realm — yoga, bodybuilding, powerlifting, martial arts, calisthenics, and CrossFit.

Yet, all of those were body awareness in motion; I lacked body awareness in stillness!

Stillness was so mentally hard for me that I had to always stuff my brain with other activities — working out, working, creating, consuming, partying — anything but the state of just being.

Thankfully, the one thing I could do was breathing well — this was the start of each meditation session and a way to get to a deeper state of interoception. I read Breathe few years ago and it forever changed my perspective on breathing. Then, Oxygen Advantage highlighted the importance of carbon dioxide tolerance and counter-intuitive shallow breathing. I’ve been regularly practicing breath work ever since, for calming effect and for performance enhancement.

Could it be that this interest in subtler bodily functions was perhaps my search for something else… and that’s when meditation came along?

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Few days in, the thought came and sent shivers down my spine: “This is it. This boiski is coming home.“

My meditation journey actually started back in 1998 and is closely connected to yoga practice that my grandma taught in Ukraine when I was just 13 years old. Grandma became somewhat of a celebrity among yogis in the post-USSR area, going from an sick math teacher at 60 years old to an incredible yoga teacher with seminars and thousands of students across the world.

Back then, I practiced yoga daily and quite seriously. I recall getting into deeper meditative states after an intense 90 min ashtanga. Teachers often talked about “Тонкий план” (astral plane in yogi lore) as something you can feel in or around your body but — being a skeptical teenager — I didn’t think much of it.

As a fantasy and DnD geek, I was also fascinated with monks and how they would train their bodies and minds to exhibit unheard of feats of strength and resistance.

Then I had to stop because my first girlfriend thought I was getting a little too removed from the real world (to be fair, a 13 year old should probably live through a full range of standard earthly emotions during that age and not get too zen about life :D).

Here I was, in Delaware, 27 years later, practicing that same technique, having similar feelings in a weird circular fashion as life often appears to be.

People say day 3 and 7th are the hardest. I found that every single day was equally hard (including 10th one).

As I was there, 10 days certainly felt way too long and unnecessary. But on on day 7 I got a profound realization and thought: “oh so that’s why we’re still here“. Then same thing happened on day 9. And even on day 10. More days means you have more chances to experience wider range of emotions stemming from different thoughts and circumstances around you, either in the now — weather, food — or in a “non-existent” mind realm (past or future).

The intended benefit of Vipassana is to learn the technique (1st-4th days), deepen it (4th-9th days), apply it in human interactions (day 10) and then take it with you to the real world via a daily practice.

The actual bouquet of benefits is much more multi-faceted. You can take advantage of each layer below individually, but together they create a powerful punch through your psyche:

1. Break from devices.
   In the modern world of social media, severing dopamine loop is extremely impactful (see Dopamine Nation, Dopamine Detox, etc.)

2. Break from civilization and daily stress.
   I noticed that without big city noise there’s a lot less anxiety. The startle reflex begets subtle bodily tension that we then carry throughout the day.

3. Undisrupted time to process thoughts.
   We rarely have dedicated time to think these days. 10-15min in a shower (look at the insights during r/showerthoughts) and perhaps a weekly 45min therapy session is all you get. Here you have hours upon hours to just think.

4. Entering meditative state that eliminates mind chatter.
   As active Beta waves diminish, they are replaced by the calmer and slower Alpha and Theta brainwaves. These are linked to creativity, enhanced learning and memory consolidation.

5. Sensory practice during meditation.
   This is the real meat of this whole experience. Who would have thought that Buddha was onto something 2500 years ago:

   1. Autonomic Nervous System Regulation: The focused attention on breath and bodily sensations sends signals to the brain that the body is safe, leading to a decrease in heart rate, blood pressure, and the production of stress hormones like cortisol.

   2. Reduced Amygdala Reactivity: By repeatedly and non-judgmentally observing sensations in Vipassana, the practitioner learns to decouple sensory experience from habitual emotional reactions, calming the amygdala's alarm system.

   3. Increased Interoception and Insula Activity: We’re now learning that interoception has a host of health benefits such as drastic improvements in anxiety disorders.

For me, Vipassana was also a challenge on other levels:

• First time no exercise for 10 days (in my 15 years of being obsessed with training)

• First time eating vegan food for 10 days (unheard of for this “foodie“)

• No supplements, no tracking your daily steps, no closing activity rings, no checking recovery score, no fasting timers, none of the other virtual shackles I’d always worry about.

Just pure existence disconnected from the rest of the world. This was the ultimate surrender and trust in the process. My ego fought this with such vengeance, it felt hard, dumb, scary, annoying, counter-productive, unnecessary, and yet… it just had to happen. I came out happier, more present, more driven by the process rather than the goal.

In 2018, I went through a stage of binging on behavioral psychology books — Thinking, Fast and Slow, Predictably Irrational, Nudge, etc. As I was listening to Goenka talk about cycles of craving and aversion and how one begets an ever-increasing loop of misery, I couldn’t help but think about all the science we know that explains the same things. The dopamine-driven reward mechanism was a great example of a reinforcing loop from Thinking In Systems and body scanning reminded me of Huberman’s NSDR (aka yoga nidra) that I’ve already been using for short rests.

I did a deep dive into the science of Vipassana teachings and wrote a separate post about it! It’s fascinating how we now have scientific backing for the entire thing.

My main critique is that there was 0 focus on movement. Walking is allowed but not encouraged. I had to stretch all the time and foam roll myself between sittings just to be able to rotate my torso. It was brutal.

I get that you’re not supposed to do anything else besides meditation and just being, but some light stretches are certainly very necessary. And the amount of sitting could probably be halfed and would still have the same effect; most people zonk out during long hourly sessions — you can’t scan your body for that long without a ton of experience.

The schedule (4am to 9pm) is also very extreme. I would argue that 4 hours of actual meditation a day is plenty, but they give you enough time to do it up to 16 hours if you really want to. For the purpose of learning the technique, it certainly doesn’t need to be so long UNLESS it’s meant to make you uncomfortable. I’m fairly certain this is one of the intentions and it certainly helps with curbing your ego.

Overall, this was the hardest thing I’ve done in a long time. And the most rewarding one. In a way, it felt like a near death experience — when you remove everything, you mind clearly understands what’s important in life. I am now more present in the moment, have a lot more clarity about what I want in life, less anxiety and the need to rush somewhere in hopes of filling my brain with activities. I have a much better sense of my body which makes it easier to not react to internal or external stimulus, especially the unpleasant one. With the pleasant ones — there’s a certain level of distancing where you still enjoy them but not in an all-encompassing, craving way. Truly an incredible tool in one’s journey of self-exploration.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

One of the things I’ve geeked out on recently was using AI to assign difficulty and performance bands to a WOD. Not just one but 907 of them (and counting).

I’m building PRzilla.app which allows you to log scores for various WOD’s, track your performance, and see all kinds of cool charts about it: how you’re progressing, what biases and weaknesses there are, movement prioritization.

In order to measure athlete performance, we need to compare it against a set of “objective” levels. For example, it is commonly accepted that Fran can be done within 3 minutes if you’re an elite CrossFit athlete, with the rest of the bands looking like this:

What is a good score for the “Fran” workout?
– Beginner: 7-9 minutes
– Intermediate: 6-7 minutes
– Advanced: 4-6 minutes
– Elite: <3 minutes

Wodwell is likely the biggest database of WOD’s and it shows bands for some of the most common ones but not all of them. I decided to add these bands to PRzilla and add them to all workouts.

But how do we measure all of them?

Ideally, we’d have a dedicated panel of experts going over thousands of WOD’s to figure all of this out. Thankfully, current top-tier AI models are trained on sufficient volume of CrossFit data and have strong-enough reasoning capabilities to do this in much shorter time.

Here’s the thing: absolute scores are bound to be subjective and context-dependent!

Even though Fran times are “commonly accepted“ as <3, 4-6, 6-7, 7-9 there can also be a decent variation among them when adjusted for male vs. female, year/decade measured (CF performance is usually trending upwards), in general population vs. experienced CrossFitters, in CrossFitters vs. specialized athletes (runners/weightlifters/calisthenic warriors), based on country/area or a specific gym (Mayhem vs. your typical box), and many more.

This makes calculations tricky but I think it’s still possible to create a range that resembles an averaged-out, close-enough representation. Our “5k run” results are likely more lax than the ones actual runners would use. But for most workouts, it’s possible to use reasoning to tell that 6 minute Fran is roughly a (top of) intermediate or a (bottom of) advanced.

More importantly, as long as our scoring system is consistent across the board, it’s great for measuring relative performance: either against yourself over time, or against others.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Top tier AI models are already trained on a large enough CrossFit data to be able to determine most of these bands. But we need few extra layers of careful orchestration to create a solid system at large:

1. SOTA thinking model(s)
   I used mostly Gemini 2.5 Pro (sometimes Sonnet 3.7) since those were the best reasoning models at the moment.

2. Prompt and context
   Make it act like a CrossFit coach / exercise scientist. “Use your extensive CrossFit knowledge”. Give it extra data to consult and base off of to narrow the scope and context, e.g. Community Cup tiers and measurements.

3. Memory bank and examples
   In practice, models are limited by 100-200K context window so we can’t send our entire JSON consisting of millions of tokens. For relative stability across batches, we need to constantly orient our model to return relatively similar calculations. I used a combination of memory bank + specific detailed documentation for this plan/feature + few examples of already existing calculations and reasoning across varied workouts. Reasoning was performed in batches to prevent hallucinations and tackle cost (this was expensive as is).

4. Internal error correction
   At the end of calculation, model needs to double-check its own analysis of specific WOD for correctness.

5. External error correction
   At the end of each batch of calculations, model takes few random existing scores and compares them to a current calculation; this ensures relative stability across many batches.

CrossFit popularized percentile-based scores during the Open, and — as part of Community Cup — they recently rolled out a scoring system that consists of 5 levels that map directly to percentiles — Rookie (<21%), Novice (22-43%), Intermediate (44-65%), Advanced (66-87%), and Pro (>88%).

I initially went with wodwell-inspired 4 tiers of Beginner, Intermediate, Advanced, Elite, then realized that there’s not enough granularity. When working on GPP charts, the scale was 1-10 and plotting “Advanced” on it was washing out the result too much. 60% and 80% could both be considered Advanced but to go from one to another might take you few years! Similarly with GPP wheel chart: if your stamina is at 60% and strength is at 80%, you would want to see that reflected on a chart as unevenness.

Here is a raw example of one of the calculations and model’s reasoning. I already had a 4 tier system/data that was generated using similar heuristic; each WOD had a difficulty, difficultyExplanation (the one model derived from its reasoning before), type, and levels.

I then used a model to derive 10 levels by giving it existing framework of how we derived those 4 levels + new system of 10 levels + examples.

Notice how it analyses WOD step by step; it understands that it’s similar to “Helen” and “Eva” since both follow a similar triplet pattern of run, x, pull-ups with this one being closer to “Eva” in terms of volume; it calculates rough times for run and thrusters while accounting for fatigue and number of rounds; adjusts edges of 10-tier to be more than 4-tier one and even considers that because it’s a “hard” WOD, beginner level is to be extended by 7min.

This now allows us to see where our scores stand for any WOD, such as this L7 that falls within 4:00-5:00 for Diane.

Before doing a workout, you can take a look at the performance guide and have a better idea which time to shoot for to get into a certain percentile.

Using similar training and reasoning, I was also able to create difficulty levels for all WOD’s. You’ve already seen “Frelen” categorized as “Hard“ earlier.

Here’s the actual documentation used when orienting AI to work with this data. AI uses this as a framework to understand general structure of a workout, and then adjusts difficulty based on modifiers like volume, skill, and load.

This produces strikingly accurate results. Note AI’s explanation for why difficulty is set certain way:

• 1k row — Easy, “A standard benchmark test of 1000 meter rowing speed.“

• Baseline1 — Easy, “A classic CrossFit introductory benchmark testing basic rowing and bodyweight movement capacity.”

• Annie 2 — Medium, “Girl WOD (Ladder Couplet). Tests double-under skill proficiency and core endurance in a fast-paced descending rep scheme (50-40-30-20-10).”

• Wittman 3 — Medium, “Hero WOD (Triplet). 7 rounds combining moderate KB swings, light power cleans, and box jumps. Tests moderate power endurance/conditioning.”

• Dork 4 — Hard, “Hero WOD (Triplet). 6 rounds combining DUs, heavy KB swings (70lb), and burpees. Tests conditioning, skill, and endurance over significant volume.”

• Kelly 5 — Hard, “Girl WOD (Triplet). 5 rounds: run, high-vol box jumps, high-vol wall balls. Tests high-volume conditioning/endurance.”

• Maggie 6 — Very Hard, “Five rounds of high-volume, high-skill gymnastics movements (HSPU, Pull-ups, Pistols). Tests advanced gymnastics capacity and endurance.”

• The Seven 7 — Very Hard, “Hero WOD. 7 rounds of 7 reps: HSPU, heavy thrusters (135lb), KTE, heavy DL (245lb), burpees, heavy KB swings (70lb), pull-ups. Extremely demanding strength/skill/volume across 7 movements.“

• Atalanta 8 — Extremely Hard, “Long Murph-style chipper with vest, high volume gymnastics.“

• 2007 Reload 9 — Extremely Hard, “Long row followed by high-skill gymnastics and heavy shoulder-to-overheads demand elite capacity and strength.“

Fun fact: “Extremely Hard“ category did not exist until I introduced Crossfit Games workouts at which point AI proactively came up with it and it made sense as the relative difficulty was objectively increased in those! Only 15 out of 907 are currently categorized as such.

Some of you will certainly scoff at a “1k row” categorized as easy. A simple movement like that can absolutely be made into a grueling test of strength, grit, endurance and stamina. The difficulty in PRzilla is not about how hard something can be made but how demanding it is on skill/strength/endurance. 1k row is easy in a sense that it can be performed by almost any person and can be completed with little effort as prescribed. You can’t say the same about Amanda that will have you do 21 ring muscle-ups together with 21 squat snatches at 135lb — feats that can take you years to master individually, not to mention being able to superset them.

Speaking of extremely hard tests, it was interesting to see how AI estimates something like “Triple unders: max reps” or “Free standing handstand push-ups: max reps”:

This is a "Very Hard" test of max unbroken triple-unders. This is an extremely high-skill movement. Even a single rep is a significant achievement for many.

High-skill jump rope variation requiring exceptional timing, coordination, and wrist speed.

- L1: Cannot complete 0 reps (effectively)
- L2: 0 reps
- L3: 0 reps
- L4: 0 reps
- L5: 1-4 reps
- L6: 5-10 reps
- L7: 11-15 reps
- L8: 16-21 reps
- L9: 22-36 reps
- L10: >=36 reps

Because we maintain relative difficulty, even an intermediate score on such tests are a great achievement. And the model understands that beginners (up until level 5) are unlikely to complete even 1.

Once we know your performance levels on all the WOD’s, it’s easy to plot them over time for a chart like this that shows “fitness level” progression and trend. And here’s something even more fun — because we have WOD’s difficulty, we can adjust your score to be more representative of real life performance (meaning that getting “Intermediate” in a “Very Hard” WOD is closer to getting “Advanced“ in “Hard” one):

adjustedLevel = cap(scoreLevel + difficultyBonus, 0, 10)

…where difficultyBonus is something simple like:

Easy: -0.5, Medium: +0.0, Hard: +0.5, Very Hard: +1.0, Extremely Hard: +1.5

Give these estimates a try — do they feel right? Could anything be improved? I’m planning to refine these in PRzilla for an even deeper understanding of workout stimulus; similar to community cup, we could be better at gender and age group adjustments. There are also gaps right now with certain WODs that have a timecap and so are a hybrid of time (if completed within timecap) and reps/load (if completed at timecap).

In the future, I’m planning to add an option to input any custom WOD and get an estimate of its difficulty and performance levels.

For Time: 500 meter Row, 40 Air Squats, 30 Sit-Ups, 20 Push-Ups, 10 Pull-Ups

50-40-30-20-10 Reps For Time: Double-Unders, Sit-Ups

7 Rounds For Time: 15 Kettlebell Swings (1.5/1 pood) , 15 Power Cleans (95/65 lb), 15 Box Jumps (24/20 in))

6 Rounds For Time: 60 Double-Unders, 30 Kettlebell Swings (1.5/1 pood), 15 Burpees

5 Rounds For Time: 400 meter Run, 30 Box Jumps (24/20 in), 30 Wall Ball Shots (20/14 lb)

5 Rounds for Time: 20 Handstand Push-Ups, 40 Pull-Ups, 60 Pistols (Alternating Legs)

7 Rounds for Time: 7 Handstand Push-Ups, 7 Thrusters (135/95 lb), 7 Knees-to-Elbows, 7 Deadlifts (245/165 lb), 7 Burpees, 7 Kettlebell Swings (2/1.5 pood), 7 Pull-Ups

For Time: 1 mile Run, 100 Handstand Push-Ups, 200 Alternating Pistols, 300 Pull-Ups, 1 mile Run. Wear a Weight Vest (20/14 lb)

For Time: 1,500 meter Row Then, 5 Rounds of: 10 Bar Muscle-Ups, 7 Shoulder-to-Overheads (235/145 lb)

Another week, another model announcement, and this time directly related to AI-assisted coding! GPT-4.1. Let’s dive in (see part1, part2, part3 for previous explorations).

First, I tried cheaper GPT-4.1 mini as a potential replacement for DeepSeek. It seemed decent at planning, but the coding was awful. It lacked agency and made way too many mistakes. Sure, it’s practically free but if you spend more time correcting and fighting with it, how much is it really costing you?

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

I quickly moved on to the “full” 4.1 and was immediately impressed by its thinking, snappiness, and general ability to get complex things done correctly. Speaking of snappy, we’re finally entering an era of talking about latency in addition to intelligence! I love that OpenAI released 4.1 with this graph:

Neither Sonnet 3.7 nor Gemini 2.5 Pro announcements ever talked about latency, yet it’s a pretty crucial aspect when iterating quickly with an AI assistant. If we compare their stats on OpenRouter (a popular aggregator used by users of Cline, Roo Code, etc.), Gemini is averaging ~10-15sec, Sonnet ~1.5sec, and GPT4.1 is at a stunning 0.6sec.

If we were to add other top models to the latency/intelligence graph, it would look something like this. To consider price difference, I’ve also added color-coded representation where green is cheap and red is expensive:

Wait, isn’t Sonnet priced higher than Gemini? Yes, but it supports prompt caching and so ends up being noticeably cheaper! Intelligence-wise, I put Claude a bit higher based on my own experience and anecdotal evidence in Discord (it’s also what powers lovable.dev) but this is subjective; I’d say Gemini 2.5 intelligence is sometimes indistinguishable to Sonnet as I rarely have issues with it when working on complex tasks so long as they’re well-defined.

So what did 4.1 fail with? It messed up a fairly simple reorganization, moving files from src/app/_components to src/app/(<route>)/components (Gemini handled it much better — it only had a couple broken tests after reorg vs. a dozen after 4.1). GPT4.1 removed a bunch of content from recentChanges.md, replacing it with <!-- ...rest of file unchanged... -->. It would often tell me what I should do instead of doing it itself or asking me to give permission to do it. And it would make basic logical mistakes like creating a map of keys to match, then lowercasing them during match, thereby causing the match to always fail.

Shortly after, I came across a GPT-4.1 Prompting Guide and it revealed a very important detail: GPT-4.1 needs to be prompted to act like an agent, meaning that it can:

1. Use tools

2. Plan tasks step by step (it’s not a reasoning model, inherently)

3. Execute tasks autonomously (and be eager and extensive about it)

The first 3 paragraphs in this recommended prompt essentially instruct that.

<!-- GPT4.1 specific instructions -->

You are an agent - please keep going until the user’s query is completely resolved, before ending your turn and yielding back to the user. Only terminate your turn when you are sure that the problem is solved.

If you are not sure about file content or codebase structure pertaining to the user’s request, use your tools to read files and gather the relevant information: do NOT guess or make up an answer.

You MUST plan extensively before each function call, and reflect extensively on the outcomes of the previous function calls. DO NOT do this entire process by making function calls only, as this can impair your ability to solve the problem and think insightfully.

Your thinking should be thorough and so it's fine if it's very long. You can think step by step before and after each action you decide to take.

You MUST iterate and keep going until the problem is solved.

Only terminate your turn when you are sure that the problem is solved. Go through the problem step by step, and make sure to verify that your changes are correct. NEVER end your turn without having solved the problem, and when you say you are going to make a tool call, make sure you ACTUALLY make the tool call, instead of ending your turn.

Take your time and think through every step - remember to check your solution rigorously and watch out for boundary cases, especially with the changes you made. Your solution must be perfect. If not, continue working on it. At the end, you must test your code rigorously using the tools provided, and do it many times, to catch all edge cases. If it is not robust, iterate more and make it perfect. Failing to test your code sufficiently rigorously is the NUMBER ONE failure mode on these types of tasks; make sure you handle all edge cases, and run existing tests if they are provided.

Once I added this to .clinerules, things have gotten much better. This means that overall experience with a model depends on the 4 pillars:

• How cheap it is

• How fast it is

• How smart it is

• How agentic it is

This distinction is important because we often see people complain about models behavior that’s actually just a result of incorrect or insufficient prompting, as I’ve experienced with GPT 4.1. On the other hand, OpenAI mentions that “GPT-4.1 is trained to respond very closely to both user instructions and system prompts in the agentic setting.” so the models themselves can inherently be more agentic than others.

Once I saw good results with my .clinerules prompt, I even started trying 4.1-mini for some tasks. Despite the instructions, unfortunately it still lacked agency, thorough step-by-step thinking through a problem, and being able to resolve it. What it was good it were small things like fixing no-unused-vars or other simple linting issues.

I think we’ll soon see a rise of tooling around a more efficient use of agents and models. I’d love to see a distributed, parallelized workflow where:

Smarter model (e.g. GPT 4.1) is creating very small, very granular tasks
for simpler models (e.g. GPT 4.1 nano)

This could improve costs dramatically. We can also imagine a world where context window is managed automatically once it reaches a certain threshold. The agent can store current progress in a memory bank, reset everything, and resume from scratch, only keeping relevant information at hand. Cline is already thinking about this: they just released /new_task command that lays foundation for it. I found that it needs some tweaking to be effective but it’s a promising start.

Speaking of context window, perhaps one of the most important charts that was shown in 4.1 announcement is how the model accuracy degrades with context window. Announcing “1M supported context window“ means nothing if your performance is at 60% after just 128k input tokens!

Other models also start to hallucinate long before reaching a context window, as shown in RULER: What’s the Real Context Size of Your Long-Context Language Models? Memory banks are important not just for cost reason but for performance as well.

This week also marks a month of working with Cline and my app is now ready! I’ve done some heavy refactoring — switched to BetterAuth which involved database changes. This wasn’t doable with vibe coding but Cline still wrote all of the code with heavy assistance. You can now edit/delete/log scores and the UI correctly changes for time vs. rep vs. load -based ones. The app has light and dark modes. It works on mobile by presenting a very different list-based view:

I was able to quickly add import/export features, and even created adjusted performance levels based on workout difficulties, which themselves were inferred with AI :)

The 70% problem that talked about is true. AI is making me extremely productive but there are many times I have to step in and ensure things go on course. One day I spent few hours trying to vibe my way through a popover not having correct styles. No amount of prompting helped and I was getting frustrated and confused, burning through more and more tokens. Then I opened DevTools and started digging through classes, quickly noticing that they take no effect. What?! Then I saw that an element was rendered at the end of a <body>. By intuition, I immediately dragged it into a main container… lo and behold, the styles applied as it was now part of a theme “context”.

Of course it’s just a matter of time until agents can use DevTools to perform the same debugging and find the same issue. We’re not quite there yet but we’re quickly approaching it.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Welcome to the 3rd installment (part 1, part 2) of assessing GenAI engineering vibe coding with Cline and top AI models.

The overall theme of the last week was: a bumpy ride but manageable with the right tooling.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

In vibe coding the right way, is not quite ready to give up on GenAI assistants. I'm right there with him—even as I'm finding more points where AI fails, I'm simultaneously discovering drastically different workflows that lead to better compliance (more on that in “memory bank“ below).

Tasks that were easy: adding basic Discord auth, table search with highlighting, UI performance tweaks. The performance challenge specifically involved a table of ~800 entries with noticeable lag during sorting/filtering (INP almost 1000ms). Gemini 2.5 Pro initially suggested memoizing sorted arrays, which barely helped. When I hinted that rendering was likely the bottleneck, not JS sorting, it quickly pivoted to extracting inlined components and preventing unnecessary re-renders—making a huge difference. I was disappointed it didn't suggest component caching first, but it redeemed itself minutes later by recommending table virtualization to render only visible chunks at a time.

It’s your choice to use AI as your minion or as a learning partner

The next big task was migrating from a simple JSON file with combined workout and score data to a database with separate tables. This took hours, not minutes. AI struggles when moving from self-contained components to systems spanning client/server. Though it mostly fixed corrupted migrations itself, I still needed to manually piece things together—setting up Turso, securing production data transfers, configuring Vercel keys—with several rounds of error log troubleshooting before everything worked.

I watched Theo’s stream the other day where he tasked various full-stack AI services with creating a party planner app. Could any of them wire backend so that there’s a working auth (via sign-in/sign-up forms)? All of them — lovable.dev, bolt.new, firebase.studio, and v0.dev — failed spectacularly. You’d have these beautiful pages and absolutely no awareness of how an entire system functions as a unit:

My experience was all too similar to his. It’s exciting that there’s so much exploration in this space but we’re just not there yet. My prediction is that with enough training, these full-stack studios will eventually start producing more coherent solutions.

Alright, on to the most important discovery. Memory bank is a neat trick folks came up with to create persistence in your AI model. Kinda like how servers use cookies to identify users via sessions since HTTP requests have no memory or association with each other. Not only that, but memory bank allows to:

• avoid ever-growing (and ever-expensive) context windows!

• create meta-level information about your project that either:

  • can’t be inferred at all (e.g. goal of the project, past architectural decisions)

  • can’t be inferred without reading a large graph of files

I was skeptical at first — duplicating info about a project that can be “easily” derived is a recipe for disaster! Turns out AI models are just like us; would you rather read a one-pager / README of a giant project or manually dig through entire codebase trying to piece how everything fits together? AI is certainly capable of doing the latter, but we live in the world where tokens cost money, and sometimes your liver.

This means following a diligent process of always reading from and always writing to your memory bank before and after each task. Here's why it matters: I removed a "date" column in the UI (the task either became too costly or got stuck). We made sure to document this change in the memory bank. Later, when fixing broken tests, without consulting that memory, AI might've wrongly added the date column back—which almost happened when one model skipped reading the memory bank. Once I reminded it to check, it immediately realized the tests needed fixing, not the UI.

Since I started using memory bank, AI is a lot less likely to get confused. I only wish I started earlier, ideally from the first commit. It also helps curb token costs since you can break your tasks into small chunks that are all done from scratch. I noticed that even cheaper AI models perform better as they have a clearer picture of your project and all the changes/decisions that happened in the past.

The downside of memory bank is extra cost upfront for each of your smaller tasks. Cline needs to send and understand each of the techContext.md ($0.0148 on Claude), systemPatterns.md ($0.0211 on Claude), activeContext.md ($0.0369 on Claude), and so on with increasing cost as the context window grows. The idea is that the final cost of completing a task is less than it would be if you hadn’t provided —and paid for—that context.

Similar tools I’ve come across: this task master for Cursor which breaks massive PRD’s into small, concise tasks.

Now that Gemini 2.5 Pro is out of experimental phase…

I accidentally $50+ in the first day of its use. Constant `replace_in_file` issues were surely one of the reasons. Gemini is still one of the best models for Cline but, while technically twice cheaper than Sonnet, anecdotally it somehow tends to rack up way more tokens way quicker, resulting in higher daily usage (~$20-30 on Claude vs ~$50-60 on Gemini). This could be lack of caching.

In other news, switching back to Claude 3.7 made me realize how slow it is comparing to Gemini. I think Anthropic’s infra isn’t great and you can certainly feel it.

I also gave Quasar Alpha a shot and it was somewhere between Gemini 2.0 and Deepseek in its capabilities — not great. It is free, though, but I don’t imagine myself using it when DeepSeek can perform better and is almost free.

A surprisingly decent free model was deepseek-chat-v3-0324 (via OpenRouter). I’ll need to experiment with it a bit more:

[…] demonstrates notable improvements over its predecessor, DeepSeek-V3, in several key aspects.

It didn’t make many mistakes but it was quite slow to respond and perform tasks.

I still find myself struggling a lot with mapping prompts to good UI outcomes. I’m not sure what Claude was “thinking” when it spit out a ScoresHistory dialog like this. It’s as if it had no understanding of any of the basic concepts of design — spacing, text size, contrast… separation of entities. Perhaps I need to look through some of these.

Out of curiosity, I fed this image to a standalone ChatGPT and it was able to create a decent looking component right there in the prompt, using React, tailwind and RadixUI:

In the next post I’ll update you on how things progress with other key tasks: switching to BetterAuth, adding score logging, and potentially exploring responsive/mobile-friendly design.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

There’s been a lot of excitement in the “vibe coding” space ever since Google released Gemini 2.5 Pro Experimental model. Last week I switched entirely from Sonnet 3.7 to Gemini 2.5 to see how it fares with building my app.

The premise is that it can perform better or at least the same and it is currently free. In my last post I mentioned price as being a big hurdle for AI coding. Naturally, a free model that performs just as well would be worth its weight in gold, especially when the alternative could cost $20-$50 daily without careful monitoring.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

TLDR: Gemini 2.5 is very promising and oftentimes indeed performs on par with Sonnet. However, it is far from perfect, both in Cline integration and in general. Despite the issues, my app development velocity increased even more, allowing me to accomplish in mere hours what would normally have consumed days of work. Price was a big factor in getting more things done quicker by re-trying failed tasks or using workarounds.

The biggest issue with Gemini ↔ Cline integration is its inability to interact with the computer environment like Claude can. It requires more manual intervention for complex tasks like fixing tests or build errors. I found myself having to npm test | pbcopy results back and forth to resolve issues, and it couldn't identify UI problems without me providing screenshots. However, when given visual context through screenshots, it performed exceptionally well at fixing issues using the available code context.

Sometimes Gemini felt like it was smarter than Claude — I loved how one time when it encountered TypeScript errors while fixing build errors, it decided to revisit them later and focus on fixing build errors first since that was a “primary objective“. This approach of tackling critical issues first before diving deeper—like a breadth-first search algorithm—mirrors my own problem-solving style. Claude would be much more enthusiastic to jump into the rabbit hole and often get stuck there.

On the downside, Gemini would stubbornly leave comments of its changes even after I explicitly mentioned in .clinerules to never do that.

It would also sometimes have its response corrupted (unsure if Cline issue) or return exact response twice (more time and token-consuming):

More critically, it committed such obvious errors that it reinforced the necessity of human oversight and intervention throughout the development process: we worked a bunch on JSON processing using jq and when I noticed that an entry was a dupe, AI stubbornly insisted that it’s not. I had to point out the flaw in its logic—it seemed to have analyzed console output completely backwards. Only then it recognized its mistake and started apologizing profusely:

As planned in my last post, I made more extensive use of .clinerules to address persistent issues, which notably improved the workflow. I still need to explore using the memory bank feature that I've heard praised in Discord discussions.

What's impressive is that as my app grows more complex, AI continues to effectively add features, modify UI/behavior, and fix errors at a consistent pace. In just minutes, I was able to implement workout frequency charts and tag/category breakdown visualizations:

Another few mins to add tests for all of the components. Making UI state persistent was also trivial for the model (here are all CF Open workouts sorted alphabetically in descending order). Interestingly, it struggled with the same things I typically do — it added state via routing in seconds before falling into an error maze with route mocking that took minutes to resolve. This mirrors my own experience of implementing features quickly, only to spend days wrestling with mocks and testing infrastructure.

I'm working with a large JSON file containing all workout data. For analysis tasks, this file needs to be transmitted to the model, which frequently struggled with handling it—taking extremely long to process, attempting to output the entire file to console, and occasionally crashing the extension entirely. It had no awareness of these issues. After pulling my hair out in frustration one too many times, I asked Gemini about efficiently reading JSON files without loading them into memory. It suggested using jq, essentially a sed/awk tool for JSON that’s built with C and has superior performance. I then prohibited the model from directly reading large JSON files and required it to use jq exclusively.

The improvement was dramatic—instead of transmitting entire files for backend analysis, Gemini would now send jq commands to be executed locally, then analyze only the relevant, much smaller output.

I used AI to fix AI performance issues but it was I who had to recognize them in the first place

This brings me to my final point — I find myself still doing a great amount of engineering, it's just that this engineering is now elevated to a higher level where I'm solving meta-challenges like:

• How to optimize AI model for performance (as explained above)

• How to optimize AI model for cost efficiency (e.g. batch requests, avoid requests that I can solve myself or that can be solved with cheaper models)

Perhaps eventually we'll see an efficiency agent that supervises the coding agent, continuously identifying patterns, suggesting improvements, and refining instructions. But for now, that's my responsibility. I've essentially become a systems engineer programming robots in a factory that build things more efficiently than I could alone.

Stay tuned for next installments in which I’ll look into memory bank, optimizing our app performance, and other explorations of top models.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Inspired by Addy’s post on using Cline for AI engineering I spent this week playing around with AI-assisted app building. Below are my initial thoughts on Cline, experience with various models, and AI engineering in general. I’m hoping to make this into a series of posts as I add more features to an app and experiment with new LLMs.

In short, Cline is a mediator between AI model and VSCode. The power comes from not only a full contextual awareness of your repo— a big boon on its own—but the fact that it can run commands, read console output, and even check browser for results of its work.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

Working with Cline feels a lot like something in between:

• Pair programming with another engineer letting them drive coding while you’re mostly talking and suggesting.

• Being a PM and telling your engineer very specific requirements of an app.

So how effective is it? Does it work? Is the coding now instinct and the 10x engineer dream a reality?

Turns out the answer is highly dependent on the model used. As mediator, Cline itself works really well. However, the difference between using Claude Sonnet 3.7 and Gemini 2.0 can feel like talking to a senior engineer with 10 years of experience and talking to someone who started coding last week and forgets to close a bracket of a function body.

So far I’ve tried Deepseek, Gemini, Claude for a variety of tasks and only Claude’s Sonnet feels like you can actually get things done fast and correctly. Both Gemini1 and Deepseek hallucinated so bad that component files would end up with a wildly-broken JSX. AI would then desperately try to fix things only to end up with more issues, then attempting to `git reset` and losing all the history that wasn’t pushed to remote; I had to `git reflog` a bunch of commits in panic.

When it works, it truly feels like magic. You can go from a sentence like “Let's update our UI in accordance to the changes made in wods.json (see file attached)“ to it working through a problem until the very end:

1. Now I have a good understanding […]

2. Now I need to examine the UI components […]

3. I need to update the sortWods function […]

4. I need to update the WodTimeline.tsx component […]

5. […rewires props, rewrites a handful of components, updates TS definitons…]

6. Now that I've updated all the components […] let's test the application

7. […runs server and checks output…]

8. I need to fix the TypeScript errors in […]

9. I notice there's a small issue in the formatScore function […]

Notice I said “when it works“. Two biggest issues you’ll notice are:

• AI doing the wrong thing

• AI doing the “right” thing but doing too much of it

The 2nd point can likely largely be mitigated by Cline's custom instructions. For example, Claude always ran `npm run dev` at the end of a task, wanting to ensure changes look good, even though I already had Next.js server running. When I saw localhost:3006 it was time to add a permanent instruction to .clinerules.

Similarly, when I asked Claude to add sorting of a table column, not only did it implement the usual sorting by clicking on the table header, but it also added a whole new UI element. This “made sense“ to AI because the column showed attempts as “Score (date 1) → Score (date 2) → Score (date 3)“ and I asked it to sort it by number of attempts, so “Score → Score“ would go before “Score → Score → Score“. Because we didn’t show actual number of attempts, it ended up helpfully prepending it on each line: “(3) Score → Score → Score“ or “(1) Score“.

This isn’t UI I imagined or even considered. Such surprising interpretations can be equally useful and frustrating.

You probably realize by now that there’s actually no “doing the wrong thing“ or “doing too much“. This is a problem of mapping the look and behavior we imagine to a concrete implementation in code. Using a lossless format like natural language will always leave room for deviation.

This is why prompt engineering often feels like art but two things should make development easier:

• Creating an extremely detailed description of UI/UX

• Working within carefully curated frameworks

The 2nd point is interesting in that it essentially forces AI to operate within a narrower context: if we’re using Radix UI with a dark theme, and you ask Claude to add a toggle, it’s smart enough to use SegmentedControl and it will be dark because the theme simply won’t allow it to look otherwise2.

AI can only mess things up as much as your building blocks will allow it to

The 1st point is something we’ll have to figure out with time but folks already experiment with some fascinating markdown -driven development as Reuven Cohen writes here. I started experimenting with this but it’s hard to figure out just how specific you have to be and which things can be omitted. I’m reminded of how before Agile came around, we had waterfall development; at my first software engineering job circa 2007, we were given 100-page long, detailed design documents of all states of an app. Have we come full circle?

The app I’m building isn’t very complex but the current UI still took me few days. Am I really 10x more productive? Or am I spending the same amount of time fixing AI hallucinations and struggling to figure out correct prompts?

Looking back, I haven’t written 95% of the code in the repo. Despite corrections and reverts, the speed at which I’m able to add features, iterate on UI, fix errors, and perform larger refactoring is probably 5x comparing to manual work. I spent 1-2hr each day on changes that would otherwise take me 4-5hr.

It’s a strange feeling to not have to write code, especially since those thornier problems are what brings the most joy (and tears), having spent hours solving them as you advance through stages of confusion, rage, and acceptance. While it feels empowering to be able to add features that you have gaps in understanding of, it also seems like a missed opportunity to learn them. But is the goal to learn or to create? The focus shifts heavily to the latter.

AI-assisted engineering feels like swimming a whole layer above the abstraction you’re so used to. The things I found as blockers were mostly design -related and so I spent most of the time tweaking UI and figuring out best way to present the data and interaction with that data:

In the recent commentary on how “90% of code will be written by AI in the next 3 months“, Theo argues that having intuition to figure out a complex problem among many layers of a system is something AI can’t solve. Yet, I’ve observed both Claude and Gemini 2.5 (arguably an even better model), reason through an error as a staff-level engineer would, considering multiple angles based on context that are far from obvious, attempting to solve them, then isolating issues and reasoning further based on what worked and what didn’t in previous steps, eventually narrowing down and solving them.

While root causes can be quickly identified by a human via intuition, you’re unlikely to catch up to AI with refactoring. More data is needed to determine how well this works in large, real-life apps but it’s looking promising. I explained to Claude what changes were made to the underlying JSON and it rewrote 5 non-trivial functions, changed 3 components, and updated all TS definitions in seconds rather than minutes. Now we’re talking way more than 10x speed.

The latest models certainly feel like a panacea but what’s the catch? I loaded up Claude with $20 worth of credits and ran out of all of them in just 2 days.

Simple tasks—by virtue of having large context window and often having to send large volume of output tokens—are usually somewhere between $0.5 and $1. As you’re iterating on your task, the counter keeps climbing up — $3.40… few mins later, $6.75. If you don’t pay attention, couple hours later you’re close to $20.

This isn’t a terribly lot of money, but it’s not insignificant. For comparison, DeepSeek is so cheap that you never have to worry about cost at all.

DeepSeek R1: $0.14 / million tokens for input, $0.55 / million tokens for output
Claude 3.7 Sonnet: $3 / million tokens for input, $15 / million tokens for output

Addy suggested to use DeepSeek for planning and Sonnet for acting, and it certainly helps, but even that approach can quickly rack up costs. And here’s the thing — the numbers might seem high but when it comes to saving engineer time that measures in $100+/hour, even Sonnet pricing is a… pretty affordable option.

Anthropic is likely aware of this which explains why their pricing tiers haven’t budged in recent months. We’ll see if the competition from a recent release of Gemini 2.5 drives these lower.

Stay tuned for future installments. Some things I’d like to try:

• Bridge the gap between design and code; try design mockups

• Improve interpretation with custom rules

• See how AI performs with more complex tasks (e.g. storing data in DB, adding users, authentication, user-level data input, etc.)

• See how it fares with even larger refactoring like switching to an entirely different UI component library

Gemini 2.5 (released couple days ago) is much better but more on that later!

You can often still end up with broken colors if your AI isn’t careful with using themed tokens; something that can perhaps be clarified in the global instructions

Ever since “getting into” CrossFit in late 2020, I've been trying to compete in the Open — an annual CF competition where you get to perform 3 workouts over 3 weeks with everyone else in the world performing same workouts. This answers 2 questions:

• How fit am I comparing to others?

• How fit am I comparing to me last year(s)?

2025 ended up being my best performance yet, and I’m generally happy, but it always comes with a mixed bag of emotions, obstacles new and old, and learnings. Ideally, these learnings are to be incorporated in the following years but I’ve been only mildly successful at that :)

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

I’ve registered for the Open in ‘21 and ‘22 but haven’t actually participated until ‘23. Not sure what exactly happened there but I think I just felt like I needed to be at the top of my game; that participating and not doing well would be a final and total defeat; that finishing in the bottom 50 percentile would result in an entire CrossFit community pointing and laughing at me. I registered again in ‘23, still hesitated and only performed 1st and 2nd workouts. Last year (‘24) was the first time I actually completed all 3 weeks.

The jump from 2023 to 2024 was massive, likely because of full completion, and it ignited the desire to take Open seriously. At that point I was also practicing CF consistently for quite a while so I ended up performing “well” and even qualifying for quarter finals (top 25% in the age group). 2025 goal was to perform at least as good, ideally better. I managed to do just that, but there was a catch…

Even though I got 1k closer to the top, I realized I was in the top third of all participants and wouldn’t even qualify for quarterfinals this year. How is this possible?

2025: 8431 / 25025 = 33.7%
2024: 9497 / 35168 = 27%

Naturally, the percentile depends on the length of the tail, and perhaps isn’t as representative of your performance. I ended up further down on the leaderboard because less people participated in the Open, likely due to last year fiasco.

Devil’s advocate: what if I got closer to the top simply because fewer of the “fitter“ people competed, hence pushing my score upwards? Counter argument: the score will always be relative to number of participants, level of participants (which constantly increases), and that year’s workouts; it should not be taken too serious to measure your own progress (more on that below).

Every year around the Open time the common theme is: life gets in a way. For example, one year we were invited to a birthday celebration of a good friend on a beach in Oaxaca with an all-paid-for accommodation, a once in a lifetime opportunity not to be missed and of course there were no CrossFit gyms in a 2 hour driving radius. This year I made sure to finalize my travels and have access to the gym a month before Open started. Yet, I was coming off of a 2-month long Asia trip so the training has been spotty and recovery poor. My jetlag lasted almost 2 weeks. As the Open started, I ended up upstate next to Hunter mountain realizing that this is my only chance to ski for the year, then skiing on 2 consecutive days 4-5 hours each, which threw a wrench into my 1st competition week. I had an absolute blast on the mountain though!

Expectations vs. reality is strong on this one. But it really comes down to “how much does this matter to you?“ And as with other goals in life, we often underestimate how much we’re willing to endure (and to give up) to bring our internal narratives to life.

I think this year was a good level of compromise but the skiing trip timing was unfortunate; I took it easy on my 1st try of 25.1 and I know I could have gotten a much better score if I retested on Monday. I suppose I can reframe this into an opportunity to get a better score next year :)

I see clearly now that what separates the good from the great CF athletes are not how much time you’re putting in the gym or any specific program or nutrition you’re running. It’s how well you can block out everything else in life in order to make training front and center.

Few years ago I read The Gains Lab’s essay on skill work and it completely changed my approach to training. The idea is that all other things equal (strength & work capacity), your proficiency in skill will determine your success. In simple terms, you can’t expect to be good at muscle-ups if you’ve only ever done them 25 times in life. Times here means days or sessions. 25 could mean, for example, that you’ve been doing them once a week for 25 weeks (~half a year). You’re probably decent at them at this point but certainly not proficient. If you’ve done them 100 times, however, you’re almost definitely going to be significantly better. Especially, under fatigue.

This is a funny metric, and my brain always tries to hijack it by coming up with counter-arguments:

• well, what if you’ve done 100 wall ball sessions in your life but then stopped doing them 3 years ago?

• well, what if you only do 1 session per month over the course of 10 years?

• well, what if your volume isn’t enough in each session? or set and rep scheme don’t change enough? does it matter if you always do the same variation of a movement? shouldn’t you add some negatives to increase the strength capacity in the eccentric portion?

While all of these are valid points, this is the kind of heuristic where you just have to paint with broad strokes. And if you do, it’s surprisingly powerful. Just don’t go into extremes — focus on frequency of practiced skill (1-3 times a week) and keep racking up that number of times you’ve done it. Eventually, it adds up to something magical. Something something 10,000 hour rule.

Following that logic, one of my goals last year was to simply rack up 100 muscle-up workouts. That’s it. I managed to do just that. Usually this meant practicing every week. Sometimes once, sometimes twice, sometimes a week would be skipped, sometimes I’d catch up and do it 3 times; or 2 consecutive days, it really didn’t matter. The were times where I’d come to the gym and bang out 10 in a row — wow! Then I’d come next week and could barely do doubles as if my body entirely forgot it all. I didn’t care about year-long progression and I didn’t care about temporary regressions. All I had to do was to get to 100/100 by the end of the year.

The result was kind of astonishing — muscle-ups turned from something I’d come to a bar and not know if I’d be able to do 1-2 on that day to something I was able to perform a set of 15 during Open 25.2 while under extreme fatigue from pull-ups, chest-to-bar, and thrusters where I could barely hold onto my grip.

Same thing happened with wall-walks which I got to 100 sessions last year, and they felt the easiest during 25.3 grueling chipper. The opposite was also true — because I barely practiced thrusters last year, they absolutely demolished me during the Open. I had to break a set of 18 into 6,5,4,3 with just 115lb on a bar.

As much as I like to get high scores at the Open, I have to remind myself that this is a very narrow and biased subset of overall fitness that CrossFit aspires to advance. As such, it’s easy to game by prioritizing movements likely to be included in it. This is why I wasn’t doing any running in the months leading up to the competition, or any GHD sit-ups, rope climbs, and heavy squats/deadlifts.

Devil’s advocate: are you better at the Open this year because you’re fitter or because you’re better at the movements that were in this year’s selection?

Of course true measurement of fitness is better answered by tracking performance over a much broader spectrum of CF benchmarks. But is the Open prep really different than a prep for a marathon where you emphasize long distance running and prepare your body for large volume as opposed to speed work? Or a Hyrox race where in addition to running you’re also improving strength and skill in a set of very specific movements.

I knew I was in trouble if thrusters were in the Open (as they always are). I tried practicing them frantically few times in the 2-3 weeks leading up to the competition but the results just didn’t stick. Forget unbroken 21-15-9, my HR would spike to 95% just doing 10 reps with 95lbs. Could it be that thruster procrastination is related to poor front rack mobility? They aren’t something I can just jump into but need 15min of prep and that’s a big deterrent. I need to hammer these weekly if I want better performance.

I also learned I’m bad with high-rep olympic lifts. I knew that my Grace (30 clean and jerks for time) score of 4:08 is “intermediate“ but it reared its ugly head twice in the recent practice, first during 23.1:

AMRAP in 14 minutes:

• 60 calorie Row

• 50 Toes-to-Bars

• 40 Wall Ball Shots (20/14 lb)

• 30 Cleans (135/95 lb)

• 20 Muscle-Ups

I was only mildly fatigued by the time I got to 30 cleans and figured moderate-pace singles will be done quickly. Instead, it took me the remaining 5-smth mins just to finish them, giving no time to do ring muscle-ups that I really wanted to squeeze in as I’ve been practicing them more lately. Few days later I ended up doing 25.3 which included almost identical pattern — 25 cleans under fatigue. I’m glad I did 23.1 and my body already knew what this would feel like, but I could have done much better if this was practiced regularly.

Conclusion: incorporate regular barbell cycling and high density oly lifts.

Finally, the other big improvement would be working on improving VO2 max via Zone 2 training. There’s plenty of evidence that this works; I just wish it wasn’t so boring.

Last year I did a VO2 max test where they strap a mask to your face—the real deal, very accurate—and it showed that my aerobic threshold is at 78% (138bpm) vs. 80-85% (~150bpm) that it usually is in trained athletes. It means the body doesn’t know how to stay in the lower intensity zone for long time and switches to a higher (aerobic) energy system. Because I often train HIIT, it’s able to stay in that aerobic zone for a long time. Overall, this results in an “excellent“ score of 55.4 for my age but widening aerobic band could be the key to an even better capacity.

Looking ahead, I’m confident I can keep getting better and inch even closer to the top next year. But the priority is still on improving overall fitness—and increasingly so—on recovery, mobility, and quality of movement as those are the foundational blocks of athletic performance.

Thanks for reading Juriy’s Substack! Subscribe for free to receive new posts and support my work.

This is Juriy’s Substack.

Subscribe now