Drafting SOPs from documents, audio and video

Owners, admins and editors can start an SOP from something that already exists.

Import a document

Use Import a document and pick a file. SOPCompass handles each type differently:

Claude is instructed to follow the document’s own structure (headings become sections, actions become steps), keep specifics exactly, never invent a step or value, and drop cover pages, tables of contents and headers. If a document contains no procedure, it returns the content as a single reference section and says so. It also returns a short note on anything you should check: sections it dropped, structure it had to infer, or places the document contradicted itself.

For narration, it turns first person talk (“then I usually export the report”) into instructions, drops false starts, keeps judgment calls (“don’t send this before the manager has seen it”), and orders steps by when they are done rather than when they were said. Where a word looks misheard, it only corrects it when the intent is clear.

Record a process or upload a video

Use Record a process to record your screen, or Upload a video for a Loom, Zoom or any other video file.

1. Your browser picks out still frames at the moments the screen changed (up to 24 are used).
2. If there is narration, your browser transcribes it.
3. Claude writes one step per frame, naming the buttons, fields and menus it can read on screen, and uses the narration for the reasons the screen can’t show.
4. Where it can clearly identify the control you used, it can draw a highlight box on the step’s screenshot. It is told to leave the box out when unsure, and boxes it rates as unsure are thrown away.

The original video can be kept at the top of the card (up to 500 MB) so readers can watch each step. That upload is for playback only and isn’t sent to the AI.

Transcription stays on your computer

Speech in audio files and video narration is transcribed with the open source Whisper model running in your browser. The first time, your browser downloads the model (about 60 MB) from the Hugging Face CDN and caches it. A five minute recording takes roughly a minute on a laptop. Recordings over 45 minutes need to be split.

Review before saving

Every import ends on a review screen. Edit the title, sections and steps, pick at least one department, and save. Nothing is added to your library until you do.