Process guide
How to write SOPs with AI without producing documents nobody reads
An SOP is not a document, it is a repeat of a decision you already made. The reason most AI-generated SOPs are useless is that they were written from the idea of a process rather than from a run of it. The method below always starts with a real run.
By Brooks8 min read
Step 1 — Capture one real run, badly
Do the task and narrate it into your phone, or type rough notes as you go. Do not tidy it. Every 'and then I check the thing' and 'this is where it usually breaks' is the part that makes an SOP worth having, and it is exactly the part a model cannot invent for you.
Ten messy minutes of capture beats an hour of writing from memory. Memory smooths over the branch conditions, and the branch conditions are where the process actually fails.
Step 2 — Turn the mess into a draft
Now the model does what it is good at: structure. Give it your notes and a strict output format.
Here are my raw notes from actually doing this task once: {paste}. Turn them into a standard operating procedure with: purpose in one sentence, when to run it, inputs needed before you start, numbered steps in imperative voice, decision points written as 'if X, then Y', a done-check, and a list of what commonly goes wrong. Do not add steps that are not in my notes. Where my notes are unclear, list the question instead of guessing.Step 3 — Answer its questions, then re-run the draft
The question list is the most valuable part of the output. Those are the places where your process is genuinely undefined, and answering them is the actual work of writing an SOP. Paste your answers back and ask for a clean second version.
Step 4 — Test it against the next run
Follow your own SOP the next time the task comes up and mark every place you deviate. A step you skip every time is a step that should not exist; a place you improvise is a missing decision point.
This is my SOP: {paste}. This is what actually happened when I followed it: {paste deviations}. Rewrite the SOP so it matches reality. Delete steps I skipped, add the decisions I improvised as explicit 'if/then' branches, and tell me which single step is still the most likely to break.Step 5 — Make it short enough to be used
Length is why SOPs get ignored. If it does not fit on a phone screen at a glance, nobody reads it under time pressure — including you.
Compress this SOP to fit one phone screen without losing any decision point: {paste}. Keep the numbered steps and every if/then. Move background, rationale and edge cases to a short appendix below. Use imperative voice, no adjectives.What makes an SOP actually hold
Three things, in my experience running these on my own operation:
- It was written from a run, not from an intention.
- It names the done-check, so you know when to stop.
- It has a date on it and gets corrected the first time it is wrong, rather than being quietly abandoned.
- It lives where the work happens — not in a folder you open twice a year.
Where AI should not write the SOP
Anything with a legal, financial or safety consequence. The model is fine as an editor on those, but the source has to be your actual obligation or a real regulation, pasted in — not the model's recollection of one.
Questions I get asked
- Can AI write an SOP for a process I have never run?
- It can produce a plausible template, but it will be generic and it will miss your branch conditions. Run the task once and capture it — the difference in usefulness is not small.
- How long should an SOP be?
- Short enough to follow on a phone while you are doing the task. Push background and edge cases into an appendix.
- How often should I update them?
- The first time you deviate from one. If you deviate and do not correct it, the document stops being trusted, and after that nobody opens it.
- What format works best?
- Numbered imperative steps with explicit if/then decision points and a done-check at the end. Prose SOPs do not survive contact with a busy day.
Where to go next
AI tools for solo business owners
What is worth paying for when there is nobody else on the payroll.
Open100 Done-for-You AI Workflows
100 ready-to-run workflows across sales, ops, content and admin. $12.
OpenThe AI Operating System
The full system these procedures sit inside: planning, production, selling, follow-up. $47.
OpenInsider List
Get the next guide when it's written
Occasional notes on running a one-person business with AI — new guides, new tools, and what actually worked. One click to leave.