How to Turn a Chat Log into an SOP

Anyone who just solved a job in Slack, email, or a ticket thread can turn a chat log into an SOP in 20 minutes: extract the real steps, one owner, and a stop. Run it in ChatGPT, Claude, or Gemini before the workaround evaporates and the next person has to ask the same six questions.

A chat log is a trail of exceptions, not a process. Force a path a trained person can run without the original humans. Context belongs in a brief. An SOP is verbs, objects, and what to do when a step fails.

Read what a skill prompt actually is, then lock the rules below.

How to turn a chat log into an SOP

Raw log first. Steps second. Prose never, except labels.

1. Export the log, then redact

Copy the thread in order. Before you paste into ChatGPT, Claude, or Gemini, strip secrets you would not put in a vendor tool.

Instruct: "Do not write a procedure yet. List actions in order as bullets: who, verb, object, and whether it succeeded. Mark questions as ASK. Mark workarounds as EXCEPTION. Mark anything with no owner as UNOWNED."

The model will try to summarize. A summary hides the exception that is the whole job. If the log is a pile, use the same raw-then-shape discipline as turning a messy note into a one-page brief.

2. Split the happy path from exceptions

Send the ledger back. Instruct: "Two lists. List A: the shortest path that worked. List B: exceptions and the trigger that caused them. If a step has no proof in the log, mark UNKNOWN. Do not invent a missing click."

UNKNOWN is the question you take to the operator. Six UNKNOWNs is a research list, not a publish.

3. Force a one-page shape

Now ask for the page. Hard cap:

• Title — verb plus object. "Refund a duplicate charge under $50."
• When to use — one trigger sentence.
• Steps — numbered, one verb each, one owner role each.
• If it fails — the matching exception from List B, or UNKNOWN.
• Stop — what "done" looks like, in a thing you can see: a status, a message, a file.

Tell the model: "Under 300 words. No rationale. If a tool name is not in the log, write UNKNOWN. Do not add a purpose statement."

4. One operator pass, not a committee pass

Give the draft to the person who did the job, or do the job once with the draft open. The question is not "does this sound right." The question is "did your hand do this."

Check three traps:

• A step that is actually two jobs.
• An owner that is a team name.
• A failure case the log showed and the draft omitted.

Fix those in the operator’s words. Send the marked-up page back: "Replace only the lines I changed. Do not re-expand. Do not add a background section."

Same discipline as a three-pass image revision loop: one class of error per pass, then stop.

5. Save the loop, ship the page

Publish the SOP where the work happens. Save the loop, not the instance, with the Promptcrates tools for reusable writing skills. Then stop. Screenshots wait until the steps survive one live run.

Reusable skill card

Trigger: A job just got solved in chat, and the next person will ask the same questions unless you write the path down.

Input: The redacted log in order, the job name in a verb-plus-object sentence, and the name of the operator who will test the draft.

Output: A ledger of actions / ASK / EXCEPTION / UNOWNED, then a one-page SOP: trigger, numbered steps with owners, failure cases, a visible stop.

Stop: After one operator run and one line-only revision. No rationale section. No merged second method. No screenshot pass in this skill.

Takeaways

• Ledger first. Happy path and exceptions second. SOP third. Never in one prompt.
• One verb per step. One owner role. A team name is not an owner.
• UNKNOWN is better than a click ChatGPT, Claude, or Gemini remembered from another app.
• One operator pass. Then stop. A wiki is not an SOP.