Stop Paying to Turn Training Videos Into SOPs. Claude Already Does It.

Somewhere in your business is a recording nobody has rewatched since the day it was made.

It is the Zoom call where your best technician walked a new hire through the one repair everyone gets wrong. It is the supplier's forty-minute product demo. It is the training video the manufacturer sent when you signed the contract. Somebody hit record. Nobody has hit play since.

The knowledge is not missing. It is just unwatched.

Every small business collects video it never uses again. A call with an accountant explaining a tax rule. A supplier walking through a new product line. A screen-share where the one person who understands your booking system shows someone else how it works.

None of it gets written down. Writing it down takes time, and time is the one thing nobody running a business has spare. So the recording sits in a folder. The knowledge sits inside it. And every time the question comes up again, somebody either remembers badly or asks the one person who was on the call, if that person is still around to ask.

You are not short on knowledge. You are short on someone willing to sit through forty minutes of video to find the ninety seconds that matter.

How do you turn a training video into an SOP with AI?

Most tools built for exactly this problem ask you to upload your video into their platform. Trupeer AI, Clueso, and Whale all work roughly the same way: you hand over an MP4 or a screen recording, the tool transcribes the audio, breaks the footage into key frames, and hands back a document with numbered steps and screenshots. That works. It also means a new subscription, a new place to store your videos, and one more login for your team to remember.

There is a free way to do the same job, and this week it showed up on the timeline of nearly every serious AI builder on X at once.

The skill two builders shipped days apart

On 2 August, marketing builder Corey Haines posted seven words that summed up the shift: "I made a skill that watches videos for me." His Claude Skill, called /watch-video, takes a link, YouTube, Loom, Zoom, Vimeo, or a local file, pulls the transcript, extracts the frames that matter, runs a vision pass over them, and files structured notes straight into his personal knowledge base. In under a day, the post picked up more than 2,300 bookmarks and 90,000 views. For an account that usually gets a few hundred, that is not noise. That is a nerve being hit.

He was not first, and he was not alone. Weeks earlier, AI builder Taoufik had published a walkthrough of a similar skill, built on an open-source project by fellow builder Brad Bonanno, that has since passed 56,000 views. Two builders, working independently, converging on the same idea inside the same month. That is usually the sign that a capability has quietly become good enough to matter.

How the skill actually works

A video is really just two things: a transcript and a set of frames. The transcript has always been the easy half. The frames were the hard part.

The old approach grabbed screenshots on a timer, evenly spaced through the video. On a one-hour recording, that is one frame every 36 seconds. On a ten-hour course, one frame every six minutes. Point that at a training video full of screen changes, settings menus, and product demonstrations, and the AI is effectively reading the transcript with a handful of random, disconnected snapshots taped on.

The fix, worked out through trial and error by builders like Taoufik and Brad Bonanno, swaps the timer for scene-change detection using ffmpeg. Instead of grabbing a frame every X seconds, the skill grabs a new frame every time the picture on screen actually changes. A demo that walks through five settings screens gets five frames, not whichever five seconds happened to land on the clock. For audio, the skill pulls captions straight off the platform when they exist, which is free, and falls back to a speech-to-text model like Whisper for recordings that never had captions in the first place, a Zoom call or a Loom walkthrough being the obvious cases.

Once it has watched the video, the skill asks one question: do you want this filed into your knowledge base? Say yes, and the note lands connected to everything else you have already stored there, searchable the next time anyone on your team needs it.

Isn't this just a fancier transcript tool?

No, and the difference matters more than it sounds. A transcript tells you what was said. It does not tell you what was on screen when it was said. If your supplier is showing you which setting to change, which wire goes where, or which box to tick on a form, that information lives entirely in the frame, not the words. A plain transcript tool throws all of that away. A skill built on scene detection keeps it, which is exactly the part a repair business, a clinic, or an agency actually needs.

Setting it up, step by step

1. Open Claude Code or Claude Cowork, wherever your team already works.
2. Install the open-source video-watching skill from its GitHub repository, a one-line clone Claude can run for you.
3. Make sure yt-dlp and ffmpeg are available. If they are not, the skill installs them for you the first time it runs.
4. Paste the link to the recording, a YouTube video, a Zoom recording, a Loom walkthrough, or a local file, and ask Claude to watch it and pull out the key steps.
5. Review the notes it hands back, correct anything it misread, then say yes when it asks whether to file the result into your knowledge base.

What this looks like in a real shop

Priya runs a nine-person appliance repair business outside Manchester. Six weeks ago, her most experienced technician, Sam, spent forty minutes on a Zoom call with a manufacturer's engineer, working through a compressor fault that only shows up on one specific model. Sam solved it. Nobody wrote it down. The recording sat in a shared drive folder named "calls - July".

Last week, the same fault came back on a different unit. The technician on site had never seen it before. Instead of hunting through old recordings or waiting for Sam to get off another job, someone pointed the video-watching skill at that Zoom recording. Ten minutes later, a structured note sat in the company's knowledge base: the fault, the diagnostic steps, the exact setting Sam changed, with a screenshot of the screen he was pointing at when he explained it. The technician on site searched, found it, and fixed the unit before lunch.

That same mechanic works whether the recording is a repair call, a new hire's first week of shadowing, a support call where a customer explained a problem in their own words, or a supplier's onboarding video for a product you sell. Anywhere your business already has video and nobody rewatches it, the same skill turns it into something searchable.

Paid platform or free skill: how they actually compare

The paid tools are not bad. They are built for teams that want a polished, exportable document with an AI voiceover on top. If what you actually need is a searchable answer the next time someone asks "didn't we already cover this?", the free skill built into a tool you may already be paying for does the same core job.

This is the same shift we wrote about when we covered turning your own expertise into a Claude Skill: you no longer need to write everything down by hand for it to become usable knowledge. The difference here is that the knowledge already exists. It is just sitting in a video nobody has watched twice.

What you have after this

A week from now, the last few onboarding calls and supplier demos stop being folders full of recordings and start being notes anyone can search.

A month from now, nobody on your team says "didn't we already go through this?" twice. If it was covered on a video once, it is one search away.

Six months from now, every recorded call your business has ever made is part of a knowledge base that gets a little smarter every time someone points the skill at a new one, including the new hire onboarding process we covered in building an AI agent that onboards every new hire, which gets far better once the source material is searchable instead of buried in a drive.

Same recordings. Same team. A completely different memory.