How to Create SOPs From Screen Recordings With AI

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Create SOPs From Screen Recordings With AI.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Create SOPs From Screen Recordings With AI.

Record the task once while explaining what you're doing and why, then give the AI the recording, or its transcript plus a few screenshots, and ask for numbered steps in your SOP format. Gemini accepts video directly, while Loom's AI plan and click-capture tools like Scribe and Tango automate much of it. Always test the draft on someone new.

The AI only knows what it saw and heard. Most gaps in a procedure are decisions experienced staff make silently: why you chose that paper setting, what you do when a customer's file arrives in the wrong colour mode, when you stop and phone the customer instead. If you don't say it aloud while recording, it won't be in the SOP, and nobody will notice until a new starter gets it wrong on a Friday afternoon.

Follow me on Instagram@sagnikteaches

Pick the route that suits the task

There are four practical ways to get from a screen to a written procedure. They differ in how much the tool does for you and how much judgement the task involves.

Connect on LinkedInSagnik Bhattacharya
RouteBest forWhat you getCost (list, USD)Watch-out
Narrated video uploaded to GeminiTasks with judgement calls, under an hourDraft steps with timestampsFree for up to 5 minutes of video; Google AI Pro ($19.99/month) extends to 1 hourCan misread small on-screen text
Transcript plus your screenshots, any assistantTeams already on ChatGPT, Claude or CopilotDraft steps you place screenshots intoYour existing subscriptionYou choose and insert the images
Loom Business + AITeams already recording in LoomSOP or step-by-step document with screenshots insertedAbout $24 per user a month (annual)Output needs editing into your house format
Click-capture (Scribe, Tango)Pure click paths in web or desktop appsA screenshot per click, auto-describedFree tiers; paid plans from about $13 to $35 per seatCaptures clicks, not reasons

On ChatGPT specifically: it accepts video attachments, but OpenAI's own help pages warn that its analysis may be incomplete and may not cover the whole video or interpret the audio accurately. For SOP work, the transcript route is more dependable there.

Subscribe on YouTube@codingliquids

Loom's route needs the least effort if your team already records there. Its AI workflows turn a recording into a document from templates that include SOP and step-by-step guide, and place screenshots from the video into the relevant steps. The free Starter plan caps videos at five minutes and doesn't include those AI features, so budget for the Business + AI plan for anyone who'll be producing SOPs.

A simple rule: if a mistake in the task would come from a wrong decision, use a narrated recording (the first two routes). If mistakes would come from clicking the wrong thing, click-capture is faster. Many businesses end up using both: Tango or Scribe for "how to raise a credit note in the accounts system", narrated video for "how to check a customer's artwork before it goes to press".

Plan the recording in five minutes

Good recordings are planned, briefly. Before you press record:

  • Choose one path. Record the normal case. List the two or three common exceptions on a sticky note so you remember to mention each at the point it would happen.
  • Clear the screen. Close email, chat and anything showing customer names, prices or addresses. Use a test order or a dummy customer where you can.
  • Make text readable. Zoom the application to 125% or more. AI models misread small interface text more often than large text, and so do people watching later.
  • Use the built-in recorder. Windows Snipping Tool and the macOS screenshot toolbar (Shift-Command-5) both record the screen with a microphone. Loom works too.
  • Narrate reasons, not clicks. "I'm clicking Preflight" is visible anyway. "I run preflight before anything else because it catches low-resolution images, which are our most common reprint cause" is the sentence that makes the SOP useful.

Here's an illustrative extract of a narration transcript from a six-person print shop, recorded by the prepress lead checking a customer's leaflet file:

"So the file's come in through the web-to-print portal. First thing, I open it in Acrobat and run our preflight profile, the one called Shop Digital A4. It's flagged two images under 300 dpi. That's fine if they're small logos, but this is the main photo, so I stop here. I don't fix it myself, because upscaling makes it look soft and the customer blames us. I email them using the low-res template, and I put the job on hold in the job board. If it's a repeat customer who's told us to just go ahead, there'll be a note on their account saying so, and then I carry on."

That minute of speech contains a tool name, a profile name, a threshold, a decision, a reason, an exception and a status change. A click-capture tool would have recorded "Click Tools, click Preflight" and nothing else.

Turning the recording into a draft

Whether you upload the video or paste the transcript, use a fixed prompt so every SOP comes out in the same shape. Save it in your team's prompt library.

Turn this screen recording (or transcript) into a standard operating
procedure using exactly this format:

Title | Purpose (one sentence) | When to use it | Who does it
Before you start: tools, access and files needed
Steps: numbered, one action per step, starting with a verb.
  - Put any decision as "If X, then Y. Otherwise Z."
  - Quote setting names, menu names and values exactly as spoken
    or shown. If you can't read a value clearly, write [CHECK VALUE].
  - Add [SCREENSHOT: timestamp] where an image would help.
Exceptions: bullet list of situations that change the steps.
Done when: how the person knows the task is complete.

Only include what is in the recording. If something seems to be
missing (a step skipped, a reason not given), list it under
"Questions for the author" at the end instead of guessing.

On the transcript route, the [SCREENSHOT: timestamp] markers double as a shot list. Play the recording, pause at each timestamp and take a still with your screenshot tool. It takes a few minutes and means every image matches a step, rather than being chosen by memory afterwards.

The illustrative draft from the prepress recording opened like this:

Check a customer artwork file before print. Purpose: catch file problems before they become reprints. Who: prepress lead or trained deputy.
1. Open the file from the web-to-print portal in Acrobat.
2. Run the preflight profile "Shop Digital A4".
3. If any main image is under 300 dpi, stop. Do not upscale. Email the customer using the low-res template and set the job to On Hold. If the customer account has a "proceed on low-res" note, continue to step 4. [SCREENSHOT: 0:41]
4. Check bleed is set to 5 mm on all sides. [CHECK VALUE]
...
Questions for the author: What happens after the customer replies? Who takes the job off hold?

Three fixes were needed, which is typical. Bleed is 3 mm at this shop; the model misread a blurry panel, and the [CHECK VALUE] marker is the only reason anyone looked. Steps 5 and 6 had been merged into one ("check fonts and colour") when they're separate checks with separate fixes. And the questions at the end exposed a real gap: nobody owned taking jobs off hold, so the answer ("whoever sent the email") went into the SOP and into the job board's rules.

Screenshots: which to keep and what to hide

A common mistake is one screenshot per click. That produces a 40-image document nobody scrolls through. Keep a screenshot at each decision point and wherever the screen looks confusingly similar to another screen. For the prepress SOP that meant four images: the preflight results panel, the hold status in the job board, the email template picker and the final "ready for press" folder.

Before anything is shared, check every image for customer names, addresses, prices and email addresses. Scribe's paid plans offer assisted redaction and its Enterprise plan enforces automatic redaction; Tango's Pro plan includes blurring, with automatic detection on Enterprise. For documents and images from other sources, the approach in redacting personal data with AI before sharing applies.

If you used Windows Steps Recorder for this in the past, note that Microsoft has deprecated it and shows a notice saying it will be removed in a future release. The replacements Microsoft suggests (Snipping Tool, Clipchamp, Xbox Game Bar) record video rather than numbered click steps, which is one reason click-capture tools have become popular.

The cold-run test

An SOP isn't finished until someone who has never done the task follows it without help. Give the draft to a colleague from another role, sit behind them, and say nothing. Write down every question they ask and every place they pause for more than a few seconds.

At the print shop, the office administrator ran the prepress SOP on a test file. She asked three questions: "Where is the preflight profile saved if it's not in the list?", "What does 'check the proof' mean, the PDF or the printed one?" and "Who do I tell if the customer's file is completely wrong, not just low-res?" Each became a line in the SOP. The third became a new exception with its own email template. The whole test took 18 minutes, and it's the step most businesses skip.

A cold run also tells you whether the procedure itself is sound. If the tester keeps hitting a step that only makes sense to the person who recorded it, the problem may be the process, not the writing. The ideas in documenting processes before adding AI help you tell the difference.

A print shop's first eight SOPs

Here's how the illustrative six-person print shop documented its most error-prone tasks over three weeks. The prepress lead recorded; the owner reviewed; the administrator ran the cold tests.

StageTime per SOPNotes
Planning and recording10 to 15 minutesTwo to eight minutes of video each
AI draft5 minutesGemini with the saved prompt
Author edit20 to 30 minutesFixing values, splitting merged steps, answering questions
Screenshots and redaction10 minutesThree to five images each
Cold run and fixes20 to 25 minutesOne tester per SOP

That's roughly 65 to 85 minutes per SOP, against the two and a half to three hours the prepress lead had spent on the one SOP she'd previously written by hand. Across eight SOPs, the AI drafts carried 11 [CHECK VALUE] markers (7 were wrong values), and the cold runs surfaced 19 questions. The shop paid for one month of Google AI Pro at $19.99 so the longer recordings fitted the upload limit, then cancelled it once the batch was done. The measure that mattered came later: over the next quarter, the new starter handled prepress checks alone from week three, and the shop logged two file-related reprints against its usual six or seven.

Other tasks this method suits

The same approach works well beyond print. Three illustrative examples, with the narration that makes each SOP worth having:

An IT support firm setting up a new user in its client's Microsoft 365 tenant. The clicks are well known; the value is in the reasons. Narrate which licence goes to which role, why the firm always enables multi-factor authentication before first sign-in, and what to do when the client hasn't said which shared mailboxes the person needs (ask, don't guess). Click-capture handles the admin-centre screens; the narrated video covers the decisions.

An events company issuing a partial refund on its ticketing platform. Record one refund, narrating the policy: refunds more than 14 days before the event go back to the card, closer than that become credit, and anything over a set amount needs the director's say-so. The AI will turn that into "If... then..." steps. The cold run should be done by whoever covers the inbox during peak season.

A translation agency creating a project in its translation software. Narrate the choices a project manager makes: which translation memory to attach, why a legal client's glossary is locked, how to set the word-count analysis. These are the settings that cost money when wrong, and they're invisible in a click trail.

Keeping recordings and SOPs in step

Software changes more often than procedures do, and a screenshot from last year's interface quietly undermines trust in the whole document. Give each SOP an owner, a review date and a list of triggers that force an update: a software update that moves a menu, a new supplier, a new product, or a reprint caused by a step that was followed correctly.

Store the recording link and its date at the top of each SOP. When the interface changes, re-record only the affected section and run the draft prompt on that clip alone, asking the assistant to "update steps 4 to 7 of the attached SOP to match this new recording and list what changed". Then cold-run just those steps.

Put finished SOPs where people will look during the task, not in a folder nobody opens. If you're building a searchable store that an internal assistant answers from, the structure in building a company knowledge base AI can answer from keeps SOPs findable. New starters should meet them in their first week; planning a new starter's first two weeks shows where they fit, and a staff training matrix records who has been trained on each.

Recording habits that produce useless SOPs

  • Recording at full speed in silence. Experienced staff fly through tasks. The AI gets clicks with no reasons, and half the steps happen too fast to read. Slow down and talk.
  • Real customer data on screen. It ends up in screenshots, in the video and in the AI tool's history. Use a test record, or plan redaction before sharing.
  • Three variations in one video. "Now if it was a repeat order I'd do this instead..." muddles the draft. Record the main path, then short separate clips for each variation.
  • Tiny text on a large monitor. A 4K screen recorded at native size is unreadable to the model and to people on laptops. Zoom the application up.
  • Skipping the "done when". Say what finished looks like: "the job shows Ready for Press and the file is in the press folder". Without it, people don't know when to stop checking.

If your team already has rough written notes for some tasks, you don't need to record those; the separate method in writing SOPs with AI from rough notes is quicker for them. Save screen recordings for tasks that are hard to describe and easy to get wrong.

Recording-to-SOP questions

Can I turn an old training call recording into an SOP?

Often, yes. A recorded Teams or Zoom session where someone shared their screen and explained a task is a good starting point. Use the transcript plus a few screenshots taken from the recording, and tell the AI which part of the call covers the task. Expect more editing than with a purpose-made recording, because training calls wander, get interrupted and skip steps the trainer assumed everyone knew.

Should the SOP be a video or a written document?

Keep both, but make the written version the official one. Written steps are searchable, quick to scan in the middle of a task and easy to update when one setting changes. The recording is useful for a first-time learner who wants to see the task done. Link the video at the top of the written SOP and note its date, so people know when it may be out of step.

How long should each recording be?

Aim for two to eight minutes covering one task from start to finish. Shorter recordings keep the AI's draft accurate and fit comfortably inside upload limits, such as the five-minute total for free Gemini accounts. If a task takes longer, split it at natural hand-off points, like 'file received' and 'file sent to press', and record each part separately.

Further reads

Sources: Gemini Apps Help (upload and analyse files); OpenAI help pages on video uploads; Atlassian Loom pricing and AI workflows support pages; Scribe and Tango pricing pages; Microsoft Support notice on Steps Recorder deprecation. Checked September 2026.

Want your key procedures captured properly?

On a 1:1 call we'll pick the procedures that matter most, choose the recording route that fits your tools, and set up the prompt and review routine your team will use.

Book a 1:1 call with me