The reason your SOPs don’t exist isn’t that you don’t know the process. It’s that writing them from scratch is tedious, and the version you’d grind out from memory would miss half the edge cases. This workflow inverts the order: the AI interviews you, you answer like you’re talking to a colleague over coffee, and the AI structures your answers into a real procedure. You then fact-check the draft against actual tool screens, stress-test the failure paths, and publish.
The trap to avoid first: typing “write me an SOP for closing the books” into ChatGPT and shipping whatever comes out. The model has never seen your books. It will produce a plausible, generic document that does not match how your team actually works, and the first person who follows it hits a step that doesn’t exist. The fix isn’t a better prompt — it’s making the AI extract your process before it writes anything.
TL;DR
- Don’t ask the AI to write the SOP cold. Ask it to interview you first (15 questions), then draft from your answers.
- Use a long-context model so the whole interview transcript stays in the draft: Claude Opus 4.7 or Sonnet 4.6 (1M tokens), Gemini 3.1 Pro (1M), or GPT-5.5. As of June 2026 all are available on paid tiers from $20/month.
- Dictate your answers — a 20-minute voice brain-dump captures detail you’d never type. Built-in voice in ChatGPT and Claude is free; Wispr Flow is $15/month for sharper, system-wide dictation.
- Always fact-check against the live tool. The AI doesn’t know your menu changed from “Billing” to “Subscription.”
- For pure click-through software workflows, a screen-capture tool (Scribe free / $23 mo, Tango free / $16 mo) auto-generates screenshots — pair it with the interview method for the judgment calls.
- First usable version: 90-120 minutes the first time, 60-75 once you’ve done a few.
Who this is for
Team leads documenting how things actually work, ops people who inherited an undocumented process, founders capturing what’s in their head before delegating, and hiring managers writing onboarding docs. If you’ve ever said “let me just do it, it’s faster than explaining,” this is the workflow that breaks that loop.
It is not the right tool for processes that change weekly (build self-explanatory tooling instead), or for workflows under security or compliance constraints that require direct expert sign-off — financial controls, HIPAA flows, anything where a wrong step has legal weight. A single “how do I do this once” task doesn’t need an SOP either; it needs a note.
Pick the right approach for your process
There are three honest categories of AI documentation tooling in 2026, and they solve different problems. Most real SOPs need two of them.
| Approach | Best for | Tools (June 2026) | What it can’t do |
|---|---|---|---|
| Interview + draft (this guide) | Judgment-heavy work: decisions, exceptions, “why,” cross-tool flows | Claude (Pro $20/mo), ChatGPT (Plus $20/mo), Gemini (AI Pro $19.99/mo) | Doesn’t capture screens for you — you fact-check manually |
| Click-capture + AI cleanup | Pure software click-throughs inside one app | Scribe (free / Pro $23/seat), Tango (free / Pro $16/seat) | Captures what you clicked, not why or when to deviate |
| Prompt-only generation | A rough first outline you’ll heavily edit | Any chatbot | Invents steps; never matches your real process unedited |
The mistake is treating these as either/or. For a process like “onboard a new vendor,” use the interview method for the judgment (“we only require a signed DPA above $10k spend”) and a click-capture tool for the rote portions (“here’s exactly how to add them in QuickBooks”). The interview gives you the spine; the screenshots fill the muscle.
Pick your model
The interview transcript plus the draft can run long, so context window matters — you want the whole conversation to stay live while the model writes the final structure. As of June 2026:
| Model | Context | Entry price | Notes |
|---|---|---|---|
| Claude Sonnet 4.6 / Opus 4.7 | 1M tokens | Pro $20/mo (Opus on Max $100) | Strongest at structured long-form writing; Pro bundles Claude Code + Cowork |
| Gemini 3.1 Pro | 1M tokens | Google AI Pro $19.99/mo | 1M context + Workspace integration (renamed from Gemini Advanced in early 2026) |
| GPT-5.5 | ~320 pages in-app (Plus) | Plus $20/mo | Full 1M context only on the $200 Pro tier; in-app context is tighter |
For a single SOP, any paid tier is fine. If you batch-process several SOPs in one project or feed in a long reference doc, the 1M-token tiers (Claude or Gemini) hold more without truncating your earlier answers.
Before you start
- Pick one process to document. Batching tanks quality; one focused 90-minute session beats four half-finished docs.
- Open the actual tool in another window. You’ll fact-check against live screens, not memory.
- Set the bar out loud: “Someone hired tomorrow could execute this on their first try.” That sentence drives every clarity decision.
- Decide where the finished SOP lives — Notion, Google Docs, an internal wiki — before you write, so you format for that home.
Step by step
- Ask the AI to interview you. Paste this prompt verbatim:
You're a process designer. Ask me 15 questions about how I do [process X], one at a time, starting broad and getting more specific. Wait for my answer before each next question. Don't write the SOP yet.The interview surfaces the steps you’d silently skip because they’re obvious to you and invisible to a new hire. - Answer in plain conversation. Don’t edit, don’t polish, just dump. This is where dictation pays off: turn on voice input (free in ChatGPT and Claude, or Wispr Flow at $15/month for cleaner system-wide dictation) and talk for 30 seconds per answer. Skipped details get caught by the AI’s next question — that’s the whole point.
- Ask for the draft in a fixed structure. Use this exact skeleton so every SOP in your library looks the same: purpose (1 sentence), trigger (when to run it), prerequisites and inputs, ordered steps, exceptions and edge cases, rollback or recovery, named owner, and revisit date. Tell the model:
Draft the SOP from my answers using only that structure. Don't invent steps I didn't describe.Consistency across SOPs beats perfect prose in any one. - Fact-check against the live tool. Walk every “click X” against the real screen. If the draft says “Settings then Billing” and the menu is now “Settings then Subscription,” the SOP is wrong on first use and loses all trust. This is the step no AI can do for you, and the step that separates a real SOP from a hallucination.
- Harden the weakest step. Find the most ambiguous instruction and ask:
What's the failure mode if a new hire misreads this step? Rewrite it so it can only be read one way.Repeat for every step someone could plausibly get wrong. - Walk the failure paths. Ask:
Generate 3 realistic "this didn't go as planned" scenarios for this process.Walk each one — does the SOP still hold? If not, add it to the exceptions section. Permissions errors, vendor downtime, partial data, and mid-process failures are the four that recur across almost every process. - Archive together. Save the SOP, the interview transcript, and the prompts in the same folder. The next refresh starts from the interview, not from scratch.
Test it on a real person
The draft is the editor’s copy; the SOP is validated by use. Before you call it done:
- Hand it to one person who has never done the process. Watch them attempt it without helping.
- Every place they hesitate or ask a question is a missing instruction. Don’t explain — write down where they stumbled.
- Edit the SOP against those stumble points. One real run usually exposes 3-5 gaps the AI and you both missed.
Quality checklist
- Could someone hired tomorrow execute it on their first try? If not, name what’s missing: context, a screenshot, or a specific click path.
- Are the steps in the order someone does them, or the order you remembered them? Those are usually different.
- Did you fact-check every “click X” against the live tool? One stale screenshot destroys trust in the whole document.
- Are exceptions captured, or is this only the happy path? Happy-path-only SOPs collapse on first real-world use.
- Is there a named owner with a revisit date? Anonymous SOPs rot; owned ones get refreshed when reality changes.
Reusing the workflow
- Save the 15-question interview prompt as a reusable template (a Claude Project, a custom GPT, or a saved snippet) so every new SOP starts identically.
- Keep your four standard edge-case categories — permissions, vendor downtime, partial data, mid-process errors — and run them against every SOP.
- Store all SOPs in one folder with the same structure. Searchability across SOPs is worth more than perfect formatting in any one.
- Refresh every six months, or immediately when the underlying tool changes its UI.
FAQ
- Should I just ask ChatGPT to write the SOP directly?: No. It will produce a generic, plausible document that doesn’t match your actual process, and the first person who follows it will hit a step that doesn’t exist. The interview step is what makes the output yours.
- Which model should I use?: Any paid tier works for one SOP. For batching or long reference docs, prefer a 1M-token model — Claude Sonnet 4.6/Opus 4.7 or Gemini 3.1 Pro (both 1M context as of June 2026) — so your earlier answers don’t get truncated.
- Do I need a dedicated tool like Scribe or Tango?: Only for pure software click-throughs, where they auto-generate annotated screenshots (Scribe free / Pro $23 a seat, Tango free / Pro $16 a seat as of June 2026). They can’t capture judgment or cross-tool decisions, so pair them with the interview method.
- Can AI be the SOP owner?: No. A human owner per SOP is non-negotiable. Owners refresh; anonymous SOPs rot.
- How long should an SOP be?: As long as the process is. A 2-step task is 2 steps; a 30-step one is 30. Don’t pad or compress to hit a length.
- Screenshots or video?: Screenshots for individual steps (they survive search and copy-paste), a short screen recording for the whole flow. Video alone isn’t searchable; text alone loses to UI ambiguity. Use both.