How to Write an SOP with AI (2026): Inputs, Prompt Template and a Test Before You Publish

Short answer: record or transcribe the person who does the task, paste that capture into an AI assistant with a structured prompt (roles, systems, thresholds, decision points, exceptions), edit the draft, then hand it to someone who has never done the task and watch where they get stuck. Fix those spots, name an owner and a review date, and publish.
Most guides on this topic are written by companies selling a capture tool, and they jump straight to "AI writes your SOP in seconds." The draft is not the hard part. The hard part is getting the real process out of someone's head, including the exceptions they handle without thinking, and proving the document works without its author in the room.
The SOP workflow at a glance
| Step | Input | AI's job | Your job |
|---|---|---|---|
| 1. Capture | Screen recording, call transcript or rough notes from the person who does the task | Turn the capture into a clean step list | Pick the expert and the real version of the process |
| 2. Structure | The step list plus roles, systems and thresholds | Fill a fixed SOP template | Supply the facts the capture missed |
| 3. Draft and question | The first draft | List gaps, vague words and missing decisions | Answer each gap |
| 4. Cold-read test | The revised draft | Nothing | Watch a newcomer follow it, note every stall |
| 5. Publish and own | The tested SOP | Write a change log entry | Assign an owner, version and review date |
Dupple data: Toolradar, our software directory, tracks 177 documentation tools as of October 30, 2026, and 77% of them are free, freemium or open source (Toolradar documentation data). Cost is rarely what stops a team from writing SOPs; the missing piece is usually the capture and test steps below.
Step 1: How do I capture the process for AI?
Record the expert doing the task once, start to finish, talking through what they are doing and why. A screen recording with voice is best for software tasks; a phone call transcript works for judgment-heavy work such as handling a refund dispute or onboarding a client.
Three rules make the capture useful. Record a real instance, not a demo, so the messy branches show up. Ask "what would make you stop here?" at each step, because that question surfaces the thresholds nobody writes down. And capture the systems by name: "the Zendesk macro called Refund-Partial" beats "the refund template."
If you only have notes, you can still start, though expect more gaps in step 3. Ask the expert to dictate the steps into a voice memo while doing the task, then transcribe it.
Step 2: What should the SOP prompt include?
Give the model a fixed structure and the facts it cannot guess. A prompt that says "write an SOP for refunds" returns a generic template that matches no real company. The template below forces the model to use your capture and to flag what is missing instead of inventing it.
You are writing a standard operating procedure from the capture below. Use only facts in the capture or in my notes; where something is missing, write [MISSING: what is needed] instead of guessing.
Structure:
- Purpose (one sentence: what outcome this SOP protects)
- Scope (when it applies, when it does not)
- Roles (who does each part, who approves)
- Systems and access needed (tool names, permissions)
- Steps (numbered, one action per step, starting with a verb; name the exact button, field or template)
- Decision points (if/then rules with the exact threshold, for example "refunds over $X need manager approval")
- Exceptions and escalation (what to do when a step fails, who to contact)
- Definition of done (how the person knows the task is complete)
My notes: [roles, systems, thresholds, compliance rules] Capture: [paste transcript or step list]
The [MISSING] instruction is the most important line. Assistants fill gaps with plausible text, and a plausible threshold in an SOP is worse than a blank, because people follow it.
For prompt variations by department, the AI Academy keeps a library of Claude prompts for SOPs.
Step 3: How do I make AI find the gaps in its own draft?
Ask a second question once the draft is back. The model is better at critiquing a document than at getting it right the first time.
Read this SOP as a new hire on their first day. List every step where you would not know exactly what to do, every vague word (such as "if appropriate", "as needed", "promptly"), every decision without a threshold, and every system you would not have access to. Do not rewrite the SOP.
Answer each point, then ask for a revised draft. Replace every vague word with a number, a name or a rule: "promptly" becomes "within 4 business hours", "escalate if needed" becomes "escalate to the team lead in #support-escalations if the customer asks twice." Those replacements are your call, never the model's.
Step 4: Why test the SOP with someone who did not write it?
Because the author cannot see the gaps. The person who knows the process fills in missing steps without noticing, and so does the expert who reviews it.
Hand the draft to a colleague who has never done the task, give them the access the SOP lists, and ask them to complete it without asking questions. Sit next to them (or on a call) and note every pause, every wrong click and every question they want to ask. Each stall is a missing line in the document. Feed your notes back to the assistant:
Here are the points where a tester got stuck following this SOP: [notes]. Add the missing information at each point, using only what I provide below: [answers].
In our view, one cold-read round is worth more than another pass by the author. Two rounds, with two different testers, is the standard to aim for on anything customer-facing or regulated.
Step 5: How do I keep the SOP current?
Give every SOP an owner, a version number and a review date at the top. An SOP without an owner goes stale within one tool change, and a stale SOP is followed confidently in the wrong direction.
When the process changes, record a short capture of just the changed part, paste it with the current SOP and ask the assistant for a redline: what to add, remove and reword, plus a one-line change log entry. Store SOPs where people already work, such as your wiki or knowledge base; our best knowledge base software and best AI knowledge base tools guides compare the options.
Should I use a chat assistant or an SOP capture tool?
Use a chat assistant for judgment-heavy processes and a screen-capture tool for click-by-click software tasks. Capture tools record a workflow in the browser or on the desktop and turn it into a guide with screenshots or video, which saves retyping every click.
| Tool | Free plan | Paid plan, billed yearly | Paid plan, billed monthly | Notes |
|---|---|---|---|---|
| Claude (chat assistant) | Yes | Pro $17/mo; Team Standard $20/seat/mo | Pro $20/mo; Team Standard $25/seat/mo | Team plan for 2 to 150 people |
| Scribe | Yes (Basic) | Pro Personal $25/seat/mo; Pro Team $13/seat/mo | Pro Personal $35/seat/mo; Pro Team $17/seat/mo | Pro Team starts at 5 seats |
| Tango | Yes, 5 shared workflows | Pro $22/user/mo for 1-2 users, $15 for 3+ | Pro $30/user/mo for 1-2 users, $20 for 3+ | Browser and desktop capture on Pro |
| Guidde (video guides) | Yes, up to 25 how-to videos | Pro $19/creator/mo; Business $39/creator/mo | Pro $29/creator/mo; Business $59/creator/mo | Business up to 5 creators |
Read on Claude pricing, Scribe pricing, Tango pricing and Guidde pricing on October 30, 2026.
Our view: start with the assistant you already pay for. Add a capture tool only when most of your SOPs are screen walkthroughs, because that is where screenshots save the most writing. For a wider list, see the best documentation tools and best AI documentation generators.
What does it cost a 5-person ops team?
On yearly billing, 5 seats cost $780 a year on Scribe Pro Team ($13 x 5 x 12), $900 on Tango Pro ($15 x 5 x 12) and $1,200 on Claude Team Standard ($20 x 5 x 12), by our arithmetic. Guidde Pro for 5 creators comes to $1,140 a year, and Guidde Business to $2,340.
For one person writing SOPs alone, the yearly totals by our arithmetic are $200 for Claude Pro, $228 for Guidde Pro, $264 for Tango Pro and $300 for Scribe Pro Personal. If budget is zero, all four have free plans (and our guide on getting Claude Pro for free covers the current trial routes), and our best free documentation tools guide compares what those allow.
What should stay human in an AI-written SOP?
Keep four things out of the model's hands. Approvals and thresholds (who signs off, at what amount) are business decisions. Compliance wording for regulated steps (privacy, financial controls, health and safety) needs the person accountable for it to write or approve it. Exceptions come from people who have handled them. And the final sign-off belongs to the SOP's owner, with their name on it.
Watch what you paste, too. Customer names, credentials and internal system URLs do not need to be in the capture; replace them with placeholders before the transcript goes into an assistant, and use your company's business plan for anything sensitive.
FAQ
Can ChatGPT or Claude write an SOP?
Yes, if you give it the real process. Paste a transcript or step list from the person who does the task into a structured prompt, and tell the model to mark missing facts instead of guessing. Without that input, any assistant returns a generic template.
What is the best prompt to write an SOP with AI?
A prompt that fixes the structure (purpose, scope, roles, systems, numbered steps, decision points with thresholds, exceptions, definition of done), supplies your capture and notes, and tells the model to write [MISSING] where information is absent. The full template is in step 2 above, and the AI Academy has more SOP prompts.
How do I test an SOP?
Give it to someone who has never done the task, with only the access the SOP lists, and watch them complete it without asking questions. Every pause or wrong click is a gap; add the missing line and test again with a second person for important processes.
Is a screen-recording tool better than ChatGPT or Claude for SOPs?
For click-by-click software tasks, a capture tool such as Scribe, Tango or Guidde saves time because it adds screenshots or video automatically. For processes that depend on judgment, a chat assistant with a good transcript produces a clearer document. Scribe Pro Team costs $13 a seat a month billed yearly, with a 5-seat minimum.
How long should an SOP be?
As long as the task needs and no longer: one action per step, decision rules with numbers, and nothing the reader does not need to finish the task. If a step needs a paragraph, it is probably two steps or a separate SOP.
Do AI-written SOPs need human review?
Yes. Approvals, thresholds, compliance wording and exceptions need a named owner to write or approve them, and the cold-read test needs a real person. The AI drafts and critiques; people decide and sign off.
How we checked
The workflow on this page is editorial guidance on structuring and testing SOPs; Dupple has no test lab and makes no timing or accuracy claims for any tool. Tool prices and plan limits come from each vendor's own pricing page, read October 30, 2026, and the team totals are our arithmetic on those printed prices. See how we rate tools; we re-check prices about every 60 days. For more ways to fold AI into daily work, see how to use AI to be more productive and how to use AI to automate tasks; our best AI for writing guide compares the assistants themselves.
Cite this: Dupple, "How to Write an SOP with AI (2026): Inputs, Prompt Template and a Test Before You Publish", November 2026.
Want to get better at AI for everyday work? Start the Dupple X trial for practical AI courses and tool guides.
Some offers on this page may be paid placements or contain affiliate links.
Across our whole catalogue the median is $23.50 a month and 37% of tools advertise a free tier. Figures are the lowest recurring monthly price quoted in our own review of each tool, read on 2026-10-23. Full method and every category.