How I Run a Martial Arts School With AI

I own Forge Krav Maga, a self-defense school in San Francisco. I've trained for about 15 years across four systems, Krav Maga, BJJ, Dutch kickboxing, and Pekiti Tirsia Kali, and I've been formally certified and teaching for ten. At Forge I teach both Krav Maga and Kali. I also have a day job leading marketing at a software company, where for the past two years I've been leading an AI transformation across a 100-person organization. And I live in the city where this technology is being built. Recently it occurred to me that I should be leaning in and finding out what AI could help me accomplish at Forge.

Most of what's written about AI for martial arts schools comes from software vendors with something to sell, and it's mostly about busywork: captions, billing emails, follow-up reminders, retention flags. That work is real, I automate some of it myself, and if it buys you back a few hours a week, good. But I think it's the least interesting part of the story. The interesting question isn't what your school gets faster at. It's what your school gets better at, and what you now have the horsepower to pursue that you didn't before.

Two things made that possible for me. One is a shift in what these tools are good at: AI doesn't hand me answers, it hands me a thoughtful, pressure-tested perspective I can argue with. The other is newer and matters more: persistent context, meaning the system I work with knows my curriculum, my teaching history, and my standards. I'll come back to that in part three, because I think it's the piece most people are missing.

One more thing up front. This has been fun. Genuinely satisfying, honestly exciting, and more helpful than I expected. That matters practically, not just emotionally: I run Forge next to a full-time day job, so efficiency matters almost as much as effectiveness, and none of this survives a bad month unless I actually enjoy the work.

The short version, for anyone skimming: I use AI at Forge for curriculum development, lesson planning, level test design, coach development, marketing strategy, member data analysis, design and video production, and editing. The judgment stays mine, and the coaching stays human. Whether you run a martial arts school, a gym, or a dojo, I think most of the patterns below should transfer, and I've tried to describe them in enough detail to copy.

Part one: the craft

Using AI for curriculum development

Forge runs a ten-level Krav Maga self-defense curriculum. It sits next to our BJJ, Kali, and kickboxing programs and borrows from them deliberately. Over the past two years I've been revising it into what I think of as modern Krav: keeping the Krav core and integrating grappling and weapons work where the legacy system had holes. Our legacy curriculum, for example, didn't teach effective side control escapes until the advanced levels and didn't teach the rear naked choke at all.

What we actually did, if you want a path to follow: digitize every level document. Load them into one project alongside the systems worth comparing against, in our case our own BJJ and Kali program materials and peer curricula. Then interrogate the whole thing structurally. Where does level three depend on a skill we don't build until level five? Where do we solve the same problem twice under two names? What can't a student do if they follow the documents exactly? Draft revisions against those findings, argue about them with your coaching team, and roll the changes out through lesson plans so they reach the mat instead of dying in a document. I didn't make these decisions alone, and I wouldn't. The coaching team made them. What AI changed is that we made them against the whole system at once instead of whatever slice we could hold in view that month. It also changed how much I can handle at speed: a dozen inputs at once, patterns found across them, changes made fast with fewer errors and gaps.

AI lesson planning from 1,500 archived plans

This one surprised me. I exported over 1,500 lesson plans, every class I've planned since I started teaching, and made the archive part of the system's working context. Then came the part worth copying: I wrote a short brief that encodes how I actually plan. It covers one theme per class, technique blocks that ladder by level instead of splitting the room, deliberately over-planned menus with explicit cut lines so a 55-minute class fits in 55 minutes, and rules for borrowing from the archive, because some material fell out of old plans by accident and some was retired on purpose, and those need opposite treatment. ‍

Then I closed the loop. Generate a plan, teach it, note what broke, update the brief. After a handful of cycles the drafts stopped making the mistakes I'd already corrected. A weekly plan now starts from a draft that follows my own conventions and mines my own decade of teaching. I tested the current version with one of my coaches yesterday. He gave Claude a two-sentence prompt, by voice, on my phone, and got back a detailed lesson plan that was probably 95% right. It blew him away.‍ ‍

Designing level tests with AI

We run quarterly level tests, and the current cycle is the most ambitious one I've built. I rebuilt our graduate-level format around less memorization and more teaching, more scenarios, and more discussion, with guest specialists the candidates have never trained with: boxers, kickboxers, Krav practitioners, BJJ players, stick fighters. The logistics are real, too. Three of my five instructors are testing themselves, and five other levels have to test with two other instructors, all in the same afternoon. Designing that is an event-production problem. For the graduate test alone it meant a 300-minute run sheet, briefs for five guest instructors, scenario cards, and gear logistics. I brought the organizing framework, roughly a quarter teaching, sparring in four modes, and situational drills carrying most of the weight, and used AI to turn the framework into documents, catch the timing conflicts, and stress-test the schedule against how tired three candidates will be in hour four. During the test I capture feedback as voice notes, and afterward the system organizes them into per-candidate write-ups, so everyone gets detailed feedback while the day is still fresh. This would have been really hard to do on my own.

AI for coach and instructor development ‍

Forge has roughly eight coaches in training, and building that program is long-form work: what a coach needs to know at each stage, how they progress, what gets evaluated. The hard part wasn't the drafting. We broke with the legacy KMG and IKMF educational model that most historical Krav Maga instructor development follows, and shifted strategies, and that was a genuinely difficult thing to do. This is where long-form drafting quality started to matter more than anything else, and it's the work that pushed me to change tools, which I'll get to. The next round is comparative, analyzing how other schools and systems structure instructor development, and I keep a running log of ideas as they surface so next year's revision starts from a list instead of a blank page.‍ ‍

Part two: the business

AI for marketing strategy

Forge always had a marketing strategy, but it was immature, underdeveloped, fragmented across my head and a dozen docs, and connected to nothing. There was no context for an AI to work with, which meant every conversation about marketing started from zero. The fix was building the strategy as a document AI could hold: an operating brief with the mission, objectives, competitive picture, funnel, and a list of settled decisions so I stop relitigating things I already decided. The brief also records how confident I am in each piece. Our core thesis is that the San Francisco Krav market is finite, three schools and limited search volume, which if true means growth runs through conversion and retention rather than reach. I think that's right, but I can't prove it yet, and the document says so.

That approach has spread well past marketing, and it's how I now handle almost any recurring decision: build a framework once, in writing, with the criteria and the confidence levels explicit, then let the AI apply it with me each time instead of reconstructing my reasoning from scratch. I don't always go with the suggestion, and sometimes I have to rebuild the framework itself. But it's almost always a better starting place: more rational, and less prone to my own biases.

Member data, MindBody, and AI

This is the least glamorous section and probably the highest-value one. Forge runs on MindBody, and there's no official AI connector, so the workflow is blunt: export the reports, screenshot the dashboards the exports don't cover, upload all of it, and ask questions in plain English. Once the data lives in the project's context, it stays queryable. "How many trials in March, and how many converted?" is now a sentence, not a reporting session.

The honest version is that I was often wrong early, and so was the AI. First-pass conclusions from clean-looking reports were confidently incorrect, and they got corrected only as we accumulated enough data, color commentary, and history from a hundred angles for the interpretations to survive contact. With that mass in place, the work got real. Cohort analysis on raw exports showed roughly half our paying members sitting on legacy pricing below the current rate. It showed trial-to-membership conversion running in the low twenties in January and February and falling to single digits by spring, and that the drop-off happens after people are on the mat, not before, which points the work at the trial experience rather than at advertising. It also surfaced the embarrassing stuff, like a referral field sitting empty on essentially every client record. The standing rule this taught me: hold the numbers firmly and the reasoning loosely.

Design, video, and social content

The same working pattern applies in a visual medium. Claude helped me build the template for both the look and the on-screen text of our reels, color-grade presets, per-clip overrides, text standards, all stored as standing context, and that saves time on every reel, every day. It also walks me through editing software I'm not expert in. Posters get built inside an established brand system, and the loop is unglamorous: generate a mockup, look at it, ask "is this right?", mark up what isn't, go again. The iteration is fast enough that I can afford taste. None of this is sophisticated. It's consistency at a volume, four to five reels a week plus event assets, that one person doesn't otherwise sustain next to a day job.

The blog: choosing what to write with AI

The writing itself is mine, and the subjects are whatever I actually want to argue about: my experiences, my perspectives, occasionally my feelings about where Krav Maga is headed. What AI changed is the selection and the speed. I keep a running list of ideas, and we evaluate them against a goal set that's genuinely complex: new members, brand, search placement, competitive positioning, what the paid keyword data says people are asking. The evaluation is a framework like any other. And then sometimes I override it, because "I want to write this" is a legitimate input that outranks the raw logic. This article scored poorly on near-term search value. I'm writing it anyway.

Teaching the AI to write in my voice

Here's the part where I'm supposed to tell you AI writes my content. It mostly doesn't, and the more interesting thing is what I built instead. Over months of drafting, I noticed I was correcting the same mistakes repeatedly: oversold claims, constructions I'd never use, a rhythm that wasn't mine. So I had the system audit my own edits, dozens of them across blog drafts and Reddit replies, and turn the corrections into a standing style document it loads before touching anything in my voice. Now drafts come back needing a pass instead of a rewrite. AI works as an editor here. The arguments have to be mine or the whole thing is pointless.

Part three: the system

Everything above depends on one piece of infrastructure, and it's the piece I think most school owners are missing when they bounce off these tools: persistent context. ‍

Persistent context and AI projects

A year ago I could not have done this. The technology wasn't ready, the interfaces weren't ready, and honestly, I didn't understand it well enough yet (thank you, day job, for the education at scale). Every AI conversation started from zero. You explained your school, your program, your constraints, got a generic answer wearing your vocabulary, and next session you did it all again. However smart the model was, it forgot you completely between sessions, and that put a hard cap on how useful it could be.

What changed is that the tools now hold standing context. Forge runs as a set of projects, one per workstream: curriculum, lesson planning, marketing strategy, the blog, MindBody data, social production, the website, paid media. Each holds its own documents, decisions, and standards. The curriculum project holds ten level docs and our BJJ and Kali materials. The lesson-planning project holds 1,500 plans and the brief that encodes how I plan. The marketing project holds the operating brief with every settled decision. When I open one, the system already knows where we left off, what we decided, and what I'd never do.

Where to start: briefs and logs

Two document types do most of the work. A brief is a standing document that encodes how you do something: your conventions, your standards, your settled decisions. It loads as context before any work begins. A log is a running capture (ideas, corrections, things that broke) that feeds the next revision of the brief. I started with one brief. Now every workstream has one, and the logs keep them honest. And I can jot a thought down anywhere, mid-day, mid-commute, and know it won't get lost. The AI won't let me forget it. ‍

Parallel chats and working by voice

Then there's parallelism, which is the part that feels like the technology finally keeping up with my brain. I run multiple chats at once, in separate rooms, none contaminating the others. I get a notification when long work finishes. I talk to Claude by voice while I bike to the gym. On a given evening I might have curriculum revision, a testing run sheet, and a data question all moving, each picking up where it left off. The flood of ideas that used to evaporate between sessions now lands somewhere, and I think the real-time, always-available multitasking makes me a faster and better business owner.

It's still not perfect. There isn't enough memory across chats and across projects yet, and I spend real time ferrying context between rooms that should share it. I think the frontier AI companies will solve this soon. The money, the expectations, the talent, and the need all seem to be there.

‍ ‍The deeper effect is that all of it compounds. Every document added and every correction made accumulates, and after two years the result is something very specifically ours. The system knows our curriculum's revision history, ten years of my actual teaching, our brand standards, our member data, my voice down to the constructions I won't use. Nobody else's AI knows any of that. A competitor can sign up for the same tools tomorrow and get none of this, because the accumulated context is where the value lives, and ours took two years of real work to build. It's the closest thing to institutional memory a one-person operation can own.

On tools: ChatGPT vs Claude for this work

‍It's the first question everyone asks. I started on ChatGPT and moved this work to Claude, for two reasons. The project infrastructure fit how I organize the work, but the deciding factor was long-form writing. Drafting a coaching development program or a curriculum revision runs thousands of words that have to hold an argument together, and that's where I found a quality gap I couldn't ignore. Take the recommendation for what it is, one practitioner's experience. I pay retail for both and have no affiliation with either company.

What AI can't do for a martial arts school‍ ‍

It can't feel a room. It can't watch a student's shoulders during a drill and know the pace is wrong. It can't coach a rep, hold a pad, or build the trust that makes someone keep showing up on the weeks they don't want to. It doesn't know when a student is about to quit, though it can flag the attendance pattern that makes me ask.

It also can't own a decision. It was confidently wrong plenty of times in the work described above, and so was I, and the difference is that I'm accountable for the curriculum and it isn't. Everything it produces is an input to my judgment. The schools that get this wrong won't fail because the AI was bad, but because someone outsourced the part of the job that was theirs.

What's next

‍There's a lot I'm looking forward to, and some of it is close. The thing I want most is a direct MindBody connection, so member and trial questions get answered live from one interface instead of through exports, ideally from my phone between classes. I want shared projects with my instructors, so the coaching team works from the same context I do, and a mobile lesson-plan interface feels very close behind that. I'm planning a comparative analysis of how other schools and systems build instructors, which will feed next year's coaching program revision. And I've just started exploring agents, meaning work that runs on its own schedule instead of waiting for me to ask. The first one I want is simple: at 7am, send me the lesson plan for today's class, drafted from the brief and the archive, so I can review and tweak it in the gaps of my day and walk into the 6pm class ready. On my end, I need more discipline about efficiency, because I currently work wastefully, and a real ceiling on this approach is how much context you can afford to keep in play.

I'm genuinely excited about all of this, and I don't say that lightly. These feel like real, step-change foundations. I think 2027 is going to be epic. Who can even guess what 2028 brings?‍ ‍

This article is part of the system too. It's written to be updated, and I'll revise it every six months as the work moves. If you're reading a version dated later than the one I published, that's the machinery working.