Originally written Jun 2026

The appeal is obvious. A coach in your pocket, available at eleven o'clock at night when you're still thinking about that role where someone got under you, and you had no answer. Infinite patience. No session fee. Access to every instruction, every competition breakdown, and every grappling debate ever transcribed to text.

For someone who spent the last few years building structured prompt systems for legal AI, applying the same thinking to BJJ felt natural. I built a custom setup: my profile (forty, not particularly athletic, overheats easily, and slow); my specific problem areas; and a workflow for turning rolling problems into research. The results were useful and limited in equal measure.

The two layers

Every sport has a cognitive layer and a kinaesthetic layer. Chess is all cognitive – pattern recognition, opening theory, and endgame positions. BJJ has both. The cognitive layer is the game planning: which technique for which position, what the competition data says about submission rates, and why a particular pass works against a particular guard. The kinaesthetic layer is the body: weight distribution, timing, and the half-inch of hip movement that turns an escape attempt from a guess into a certainty.

AI lives entirely in the cognitive layer. It has never been choked.

It can describe in accurate detail exactly why you should keep your elbows in close during a back escape – but it can't see that yours aren't.

What it gets right

The game planning work has been genuinely useful. Before settling on the technique list from the previous post, I used Claude to map high-percentage techniques to my specific profile – slow, pressure-based, not relying on speed or explosive movement. That's exactly what AI is built for: who are the right instructors to study? What does competition data say about submission rates at my level? If I'm going to invest hundreds of hours drilling escapes, which ones are worth the rep budget?

The workflow runs like this: note the specific problem from a rolling session, query Claude with enough context to get a useful response (my profile, the position, and the specific failure mode), and validate the recommendation against real coaching before committing it to Mat time. The custom setup helps significantly – Claude's responses are more targeted when it knows it's talking to someone who overheats easily and won't be winning any speed contests.

For comp prep research, for understanding the "why" behind techniques, for organising drilling blocks, for identifying the right instructors for a particular position – consistently useful. This is the cognitive layer, and the cognitive layer is where it earns its keep.

What it gets wrong

The gap becomes obvious the moment you try to apply the advice without the validation step.

I was working through turtle escape sequences – a known problem in my game. The recommendations were technically accurate: the Priit Mihkelson framework, the specific positions, and the logical progression from turtle to sitting guard. What Claude couldn't know was that my posture going into the turtle was creating the problem, not my escape technique. My neck was exposed before the escape even started. That correction came from mat time and from coaching, not from a text response.

This isn't a failure of AI specifically – it's a structural limitation. Research into how language models represent knowledge shows they diverge most sharply in motor domains, the exact domain grappling requires. An LLM can accurately describe a technique and still be telling you nothing useful about how your specific body needs to execute it. The description and the movement are different things.

The hallucination risk is also real in a niche sport. The more specialised the topic, the more confidently wrong a model can be. Validating AI-generated advice with coaches before committing it to Mat Time isn't extra work – it's the whole reason the workflow produces results rather than bad habits.

The fix

The model that works is clearly hybrid. AI handles the cognitive layer: research, technique selection, game plan structure, drilling plan design, and comp prep analysis. Coaches and rolling handle the kinaesthetic layer: what my body is actually doing, where the positions are breaking down, and corrections that can only come from someone watching.

The custom instruction set matters here. It's not magic – it's making sure Claude has enough context to be useful. My profile, my constraints, and the specific rolling problem are stated precisely. The same structured approach I've applied to building prompts for legal AI systems works in a BJJ context too, with one critical difference: there's always a coach or a rolling session at the end of the process to validate whether the output was actually right.

What it changed

The technique list came partly from this process – high-percentage techniques mapped to my constraints and instructors whose systems fit a pressure-based game for an older body. AI identified the right corners of the instructional library. The mat and the coaches confirmed what actually worked.

What shifted more fundamentally was how I framed the questions. Vague queries produce vague answers. "How do I get better at escapes?" is a bad prompt. "I'm forty, slow, defaulting to turtle too often, my neck keeps getting exposed, what's the systematic framework for improving back escape defence?" is a better one. The quality of the answer depends almost entirely on the quality of the question, which is true for human coaches too, but especially true for AI.

A note on AI coaching

The version that doesn't work is the passive one – querying without context, taking advice without validating, and using it as a shortcut rather than a research layer. The version that works knows exactly what AI is and isn't for.

Off the mat, for the thinking side: useful, sometimes excellent. On the mat, for the physical side: not present, not possible, not a substitute for anyone who can actually watch you move.

AI can describe exactly how to escape from the back. It just can't tell you your elbows were already in the wrong position before you started.