AI Video Coaching Review: How I Built a Multi-Pass Critic (Simply Explained)

How a Good Coach Actually Watches Tape

Watch a great martial-arts instructor review a student's video. They don't watch it once and shout out a grade. They watch the whole clip first to get the shape of it. Then they pick the three or four moments that actually matter. Then they go back and re-watch those moments slowly, sometimes frame by frame.

That two-step attention is the whole job. The first watch is about the big picture. The second is about the details. Skip either one and the feedback is worthless.

I built a coaching tool for a martial-arts academy. Students upload a video of a technique, and the system gives them feedback. Simple on the surface.

But here's the trap. If you hand a video to a single AI and say "grade this," you get a confident-sounding paragraph that no real instructor would respect. "Good energy, work on your base." Vague. Useless. The kind of thing that makes a student feel coached without actually coaching them.

So I built it to copy what a real coach does. Watch the whole thing once to find the important moments. Re-watch those moments in detail. Then put it all together.

The Feature Was Broken Before the AI Even Ran

This is the part business owners need to hear, because it's the reason "we tried AI video and it didn't work" is so common.

The feature looked done. It had a button. It returned text. But it had never once actually analyzed a video.

Here's what was happening. The system was handing the AI a streaming link, the kind your phone uses to play a video smoothly. That link is great for watching on a screen, but it's useless to an AI. It's like handing someone a recipe card and expecting them to taste the meal. The AI got a pointer to nothing it could actually use.

And there was no error message. The AI just made up plausible-sounding feedback based on essentially nothing, and everyone assumed it worked.

That's the most expensive kind of failure. A crash gets noticed and fixed. A confident wrong answer ships and nobody catches it.

The fix was boring plumbing, which is usually where the real work lives. On upload, I have the system create a proper video file the AI can actually read, then hand that file directly to the AI. I check that it loaded correctly before doing anything else. And I upload the video once and reuse it for every step after, so I'm not paying to upload the same clip five times.

The lesson I keep relearning on client work: the difference between a working AI feature and a broken one is usually the delivery, not the cleverness of the prompt. Most people spend a week fixing the wrong thing when the AI was never getting a video it could see in the first place.

Watch Once to Triage, Then Zoom In

With the AI finally getting a real video, the actual coaching could start. I built it as a team of AI specialists, each with one job, instead of one overloaded generalist.

The first specialist watches the whole clip once at normal speed. Its only job is to break the technique into phases (setup, entry, the move itself, the finish) and pick up to four moments worth a closer look. That's it. It's not grading anything. It's the instructor scanning the clip to decide where to spend attention.

The four-moment limit is on purpose. A coach who comments on every half-second gives you noise. A coach who picks the four moments that decide whether the technique worked gives you something you can act on. The limit forces that same discipline.

Then comes the zoom step. For each of those four moments, a separate specialist re-watches just that slice of video, this time looking very closely, frame by frame. This is the difference between "your base looks off" and "at 8 seconds your weight rolls back onto your heels right before the throw."

All four zoom passes run at the same time, like four assistants each studying one clip at once, so it stays fast.

Here's the part that matters for any video AI you're thinking about. Splitting the work into focused jobs beats one giant "analyze everything" request. A narrow job gives you a specific answer. A giant request gives you mush, because the AI is trying to hold the whole clip and every possible critique in its head at once.

Narrow focus, specific answer. Wide focus, vague answer. That holds across nearly every system I build.

Put It Together, Then Let a Human Decide

The last specialist takes everything (the phases, the detailed findings from each zoom pass) and produces one result: a score plus feedback tied to exact timestamps.

The timestamps are what make it useful. "At 8 seconds your weight is back on your heels, which kills your entry" is coachable. A student can rewind to that exact moment and feel what the AI saw. "Work on your base" is a fortune cookie.

I'll be honest with you, because I'd rather you trust me than oversell. AI scoring is a strong, consistent first read. It is not a black belt with twenty years on the mat. Which is exactly why the last piece exists.

That score affects rank. So I had to fix a problem hiding in plain sight. In the first version, a student could trigger their own grade and promote their own rank. The AI was the judge, and the student held the gavel.

The fix lives on the server, not in the design. Students can upload and get draft feedback all day. But finalizing the grade that actually counts is locked to instructor accounts. Hiding a button isn't security. Checking who's allowed to push it is.

This is a principle I apply to every system I build: the AI proposes, a human disposes. Anything that moves money or moves rank stops for a person. The AI does the watching. A human owns the decision.

Why This Matters Beyond Martial Arts

If you've tried AI video review and walked away unimpressed, here's what probably happened. Someone did it in one pass, and there's a good chance the delivery was quietly broken and nobody noticed, just like mine was.

The pattern that actually works is the one I just walked through. Fix the plumbing so the AI can see the video. Watch once to find the few moments that matter. Re-watch those moments closely, in parallel. Pull it all into feedback that points at a specific moment. And keep a human in charge of any decision that counts.

None of this is specific to martial arts. Swap the subject and it holds. Sales calls where the AI flags the three moments a rep lost the deal. Factory footage where it isolates the frames a defect appears. Training, compliance, sports, anywhere a human currently watches video and writes notes by hand.

The thing I want you to take away: the design matters more than the AI. The same AI produced garbage on a broken link and genuinely useful coaching once the system around it was right. The AI didn't change. The setup did.

Want to explore what AI could do for your business?

Book a free 30-minute strategy call. No pitch deck, no sales team, just a real conversation about your operations and where AI fits.

Book a Discovery Call