Turn a Screen Recording Into an SOP and Let AI Write the Steps

Almost every business has a process that only lives in one person's head: how orders get entered into the system, how the monthly report gets built, how stock is checked before goods ship. Everyone agrees it should be written down as an SOP. The problem is that nobody has time to write it, and the person who understands it best is usually the busiest.

There's a much lighter route: that person just records their screen while doing the process as usual, explaining what they're doing along the way. AI drafts the SOP. In this article I cover how to do it, which tools help, and what a human still has to check.

Why SOPs rarely get written

Writing an SOP is boring and slow. The person who knows the process has often forgotten which steps are "invisible", because they've become automatic. So when that person goes on leave or resigns, the process stalls. That problem of knowledge stuck with one person is also what I covered in AI Is Taking Over Junior Work, So Who Becomes the Senior Later.

A screen recording flips the order: the person doesn't have to write, just work and talk. Putting it into words is handed to AI.

Two approaches you can use

1. A click recorder. Tools like Scribe run as a browser extension or desktop app. You press start, do the process, then press finish, and the tool builds a step-by-step guide from your clicks, complete with screenshots. You can edit the result and cover up sensitive information in the screenshots. It suits processes that happen entirely in web-based apps.

2. A video read by an AI model. Record your screen with any recorder, then give the video to a model that can read video. According to the Gemini API documentation, Gemini can describe, segment, and extract information from video, answer questions about its content, and refer to specific timestamps. It processes the audio and the visuals together, so your spoken explanation gets read too. One important detail: by default the video is sampled at 1 frame per second. A very fast click can fall between frames, which is why talking it through and keeping a slow pace matter.

How to record so the result is good

These are my suggestions based on how the tools work, not rules from any one tool:

1. One process per recording, ideally 5 to 15 minutes. Split longer processes up.
2. Talk while you work, including the reasons. Not just "click this button" but "click this because if the status isn't paid, the invoice mustn't be sent".
3. Slow down at the important steps. Pause briefly before and after the clicks that matter.
4. Mention the exceptions. "If it's a corporate customer, the PO number is required." That's the part most often missing from SOPs.
5. Use sample data, not real customer data. A screen recording easily captures names, phone numbers, or amounts that shouldn't be stored in more places.

A prompt for drafting the SOP

Once you have the video, ask the AI to draft it in a clear format. Here's an example you can use:

You are an SOP writer. From this screen recording and my voice explanation, write an SOP in this format:
1. Purpose of the process (one sentence)
2. Who does it and when
3. What needs to be ready before starting
4. Numbered steps, one action per step. Name menus and buttons exactly as they appear on screen
5. Exceptions: if ... then ...
6. How to check the result is correct
Don't add steps or reasons that aren't in the recording. If something is unclear, write [UNCLEAR: ...] so I can check it.

That last line is the most important. Rather than letting the AI guess, it's better for it to flag what's unclear.

What a human still has to check

• Missing steps or steps in the wrong order. Especially when the clicks were fast.
• Numbers, menu names, and business rules mentioned only in passing in the recording.
• The reasons behind steps. AI can make up a reason that sounds plausible. If you didn't say the reason, don't let it appear in the SOP.
• Test it with someone else. Ask someone who has never done the process to follow it using only the SOP, without asking questions. Every time they get stuck, that's where the SOP needs fixing. I think this is the most honest check there is.

On checking AI output more generally, I go deeper in At What Point Should Humans Still Check AI and Automation Results.

Store it somewhere the company owns, and keep it updated

An SOP sitting in someone's personal folder or personal chat might as well not exist. Store it on a company-owned drive, give it a clear name, note the date it was last updated, and decide who owns it. When the process changes, re-record only the part that changed. The same principle applies to the team's prompts and AI accounts, as in A Simple SOP to Keep Company AI a Team Asset, Not Personal Property.

Wrapping up

An SOP doesn't have to start from a blank page. Record the process, explain the reasons, let AI draft it, then test it with someone who's never done it. What used to live only in one person's head now belongs to the team.

If you have a process you'd like documented or even automated, tell me about the process first.