Why Your SOP Doc Has Been 60% Done For Six Months
Why your SOP doc rots and the interview that fixes it
Surgeons asked to describe their own procedures left out 73% of decision steps. They perform the procedure every day. The half-finished SOP doc on your desk is the same retrieval gap, in a different uniform.
Surgical Cognitive Task Analysis studies measured a 73% gap. Experts asked to describe their own routine procedures left out 71% of clinical knowledge steps, 51% of action steps, and 73% of decision steps in free recall. The format of asking changed the answer.
The half-finished doc isn't a discipline problem
That SOP doc has been at 60% done since November. You blocked off an hour with Marcus three times — each one slipped. Now you're trying to write it from memory based on the ride-alongs you've done. Three Notion templates open. Nothing landing.
The problem is the prompt, not the person.
The Marcus pattern — senior person knows it, doc rots
Every ops business has a Marcus. The senior tech who runs the route nobody else can run. The closer who handles the escalation calls that go sideways. The chef who keeps line speed up when the rush hits. They know the work. They cannot easily write it down.
Not because they're lazy. Because procedural knowledge — the kind that runs ops-heavy businesses — is mostly tacit. Experts retrieve it in real-time as they act. Asking them to abstract it into a template is asking them to translate a language they think in into one they don't.
Why templates fail — they ask the doer to do two jobs at once
The Notion template asks Marcus to remember the steps AND to write them down AND to think about which ones matter to the new hire. That's three cognitive jobs stacked. Marcus does one of them well. The other two get half-finished. Operators who've run the interview instead tend to ship the SOP in a single afternoon.
The CTA research from surgery — 73% of decision steps left out
Cognitive Task Analysis is a methodology developed in the 1980s for the aviation and military to extract procedural knowledge from experts who couldn't articulate it. Researchers (Sullivan, Hwang, Park, et al., published in American Journal of Surgery) ran it on surgeons describing routine procedures. The result: surgeons left out 73% of decision steps in free recall. They perform the procedure every day. They could not describe it.
Why interviews extract what templates can't
The fix isn't more discipline. It's a different prompt. CTA studies measured what happens when researchers swap free recall for structured interviewing about a recent specific case. The articulation rate jumps from ~30% to near-100%.
Experts omitted an average of 71% of clinical knowledge steps, 51% of action steps, and 73% of decision steps. — Sullivan, Hwang, Park, et al., on Cognitive Task Analysis in surgical education
Free recall surfaces ~30%. Structured interviews surface the rest.
The mechanism is straightforward. Asking "what do you do when X happens" returns an abstract script. Asking "what did you do last Tuesday when X happened" returns the actual procedure — including the gotchas, the workarounds, and the 'done when' check the doer reads in their head but never writes in a doc.
The "recent specific case" anchor
The single most important interview move: pick a real case from this week. The Marquez install. The Tuesday callback. The escalation Saturday night. Not "the process in general." General questions get general answers. Specific cases surface tacit knowledge.
'Done when' checks emerge in conversation
The other thing interviews surface that templates miss: the signal the doer reads to know a step is finished. Marcus doesn't write "verify combustion is clean" on a template. He says it out loud when you ask him "how do you know you can move on?" That signal IS the SOP. It just lives in the conversation, not the doc.
Running the interview yourself, in one conversation
The good news: you don't need a CTA researcher. You don't need to learn the methodology. You need the build prompt at the bottom of this page, an AI account you already have, and 90 minutes with Marcus on speakerphone.
The build prompt — paste into Claude or ChatGPT
The prompt is the structured interview. You paste it into a fresh chat in Claude or ChatGPT. Tell it the process you want to document. Put Marcus on speakerphone. The agent asks the questions; you answer them one at a time on Marcus's behalf, or hand him the phone.
What good output looks like
A walked-through SOP for one process, 6-12 numbered steps, each carrying a 'done when' check and a gotcha named in Marcus's actual words. At the bottom: the tribal-knowledge anchor — the one thing Marcus has never written down because it felt too obvious to him.
When to re-run — every gotcha that costs you
The first SOP is the starting point. The doc lives — every gotcha makes it sharper. Every time a gotcha causes a callback, a refund, or an escalation, run a 15-minute follow-up interview. The agent updates the SOP. The next new hire reads it and ships the work.
The doc that's been at 60% done since November isn't broken because nobody cares. It's broken because the prompt was wrong. Change the prompt and the SOP ships this afternoon.