Record the task once. Never explain it again.

The recipe takes one input, the transcript of a narrated screen recording, and produces a complete SOP in a fixed format: purpose, owner, tools and access needed, preconditions, numbered steps, what done looks like, and a list of gaps.

The formatting template is built into the prompt, so every SOP your team produces comes out in the same shape. Each step starts with a verb and names the exact screen, button, or field the narrator mentioned. When the narrator explained why a step matters, that reason is kept as a note under the step. When they said "unless" or "if," the exception is captured too, because the exceptions are the part of tribal knowledge that never survives being written down by hand.

The gaps list is what keeps the output honest. The prompt forbids the model from inventing steps or settings. Anything the narrator skipped, mumbled, or contradicted becomes a direct question for them instead of a plausible-sounding guess.

Setup is one recording. The person who actually does the task, not the manager who thinks they know it, does a real run while saying everything out loud: clicks, keystrokes, what they are checking for, and every exception. Loom and most meeting tools produce the transcript automatically; if yours does not, most AI assistants accept the audio or video file directly.

Then paste the saved prompt and the transcript, and read the gaps list first. Answer its questions in a reply and ask the model to revise. Skim the steps against your memory of the recording and delete anything that sounds plausible but was never said.

The QA step is the part most teams skip and the part that matters most. Hand the SOP to someone who has never done the task and watch them follow it without helping. Every pause or improvisation marks a hole in the document. Fix those steps, test once more, then add an owner and a review date.

This loop is the right tool for documenting your first handful of processes. It strains when dozens of SOPs go stale every time a tool or policy changes, when nobody trusts the docs because the process has moved on, or when the real goal was never a document but getting the work off a person's plate entirely. A finished SOP is also the best possible spec for an AI agent: the steps, exceptions, and done-state are exactly what an agent needs to execute the procedure instead of a human reading it. That handoff is a build, not a prompt. See what we install or book a free AI opportunity audit to find out which of your documented processes could run themselves.