SOPs were written for people to read. Paragraphs, context and "use your judgment" suit a human. AI works better with steps.
This guide gives you one prompt that rewrites every step with four labels. The label that matters most is STOP AND ASK, because it tells Claude where to pause.
01 / 08
The four labels
Each step in the SOP gets four fields.
- ACTION. One instruction, starting with a verb. Twenty words at most.
- INPUTS. What the step needs, and where that comes from.
- CHECK. How you know the step was done right.
- STOP AND ASK. YES when a person must approve before moving on.
The first three turn a document into a to-do list. The fourth gives you a checkpoint.
02 / 08
What gets a STOP AND ASK
Flag any step that does one of these:
- Sends anything to anyone outside your team, including a welcome email or a follow-up
- Takes or spends money, including charging a client's card
- Deletes or archives something
- Needs a judgment call
At each flagged step Claude should pause and show you what it is about to do. You say yes and it carries on.
03 / 08
Run the conversion
Pick the SOP your team ignores the most, the one they open least. Paste the prompt into a new chat, with the SOP at the bottom.
04 / 08
Answer the open questions
Claude does not fill gaps with a guess. If your SOP says "update the client" without saying how, that goes under OPEN QUESTIONS.
Answer each one in the chat, then ask Claude to show the checklist again with the answers in place. Before you save, read the final checklist next to the original SOP, line by line, and check no detail was dropped. Save the finished version somewhere your team can find it.
05 / 08
Run the checklist
Open a new chat, paste the finished checklist and write the line below. Replace {TASK} with the details of this run, for example: new client Tilly Barnes, 14 Elm Road.
Watch the first run. Check that Claude really stops where it should. Without a connector, Claude cannot click anything in your tools: it tells you what to do at each step and you do it. With a connector it can act for you, which is why the next box tells you to block sending in the connector as well.
06 / 08
A worked example
Picture a small design studio with six people. Their SOP for onboarding a new client is one long paragraph that the account manager wrote two years ago. Part of it reads: "Once the deposit clears, set up the project folder, send the welcome email, and get the kickoff call booked. Chase the brand files if they haven't arrived."
The owner pastes the SOP into the prompt above. Claude gives back a slice like this:
3. ACTION: Send the welcome email to the new client.
INPUTS: Client name and email, from the signed proposal.
CHECK: The email names the right client and the kickoff date.
STOP AND ASK: YES. Show me the full email before it goes.
OPEN QUESTIONS
- "Once the deposit clears": who confirms this, and where do they look?
- "Chase the brand files": how many days of waiting counts as late?Two things happened. Step 3 sends something to a client, so Claude flagged it without being told. And the SOP's vague phrases landed under OPEN QUESTIONS instead of being guessed at.
The owner answers both questions in the chat: the bookkeeper confirms the deposit in the banking tool, and brand files are late after three working days. Then she asks Claude to show the checklist again. The answers now sit inside the steps, and she saves it in the shared team folder.
On the first run she pastes the checklist, adds the run line with the new client's details, and watches. At step 3 Claude stops and shows the draft email. She changes one sentence, says yes, and the run continues. That one edit would have gone out unread without the stop.
The next step in her checklist books the kickoff call. Her Claude chat has no connector to the calendar, so it writes the invite text and tells her who to send it to. She sends it herself. Nothing in the run depends on Claude touching her tools, which makes the first few runs a safe place to learn what the checklist gets right and where it needs another pass. After three runs she has fixed two vague steps and removed one that nobody actually did.
A useful habit after each early run: ask Claude, "Which steps did I have to correct or explain?" Each answer points to a step that needs clearer INPUTS or a sharper CHECK. Update the saved checklist right then, while the run is fresh in your mind, and tell the team which version is current.
Why this works
Claude follows instructions literally. A paragraph that says "use your judgment" gives it nothing to follow, so it fills the gap with a guess. Short numbered steps with a CHECK give it a clear job and give you something to compare against. The forced labels also expose how much of your SOP lived only in someone's head.
07 / 08
If it goes wrong
Common mistakes
- Converting an SOP nobody has read in years. If the process changed since, the checklist will faithfully copy the old process. Read it first and fix anything that is out of date before you paste it.
- Answering open questions with "it depends".
Claude will turn that into another vague step. Give a rule it can follow, such as "three working days" or "the person who signed the proposal".
- Trusting the flags without reading them. The prompt catches the four risky kinds of step, but others can be hard to undo too, such as marking a job as closed or overwriting a shared file. Read the final checklist once and ask of each step: could this be hard to undo? If yes, change it to YES.
- Never testing on a real case. A checklist that looks right on screen can still fail on the first live run. Use a low-risk task and watch every step.
- Saving the checklist and never updating it. When the process changes, change the checklist the same day. Otherwise you are back to an SOP nobody trusts.
Fixing a run
Claude skipped a STOP AND ASK.
Reply: "You skipped the stop at step 6. Start again from step 5 and stop where flagged." If it repeats, block the sending tools in the connector's tool permissions.
The checklist has more steps than my SOP.
Reply: "Remove steps the SOP does not imply." Keep the SOP's order.
My SOP is two pages long.
Split it by section and convert one section at a time. Short checklists are easier to check.
Claude merged two of my steps into one.
Reply: "Split step 4 into two steps, one action each." If a merged step has a STOP AND ASK, you cannot approve one half without the other.
Claude marked almost every step STOP AND ASK.
That is safer than too few. Reply: "Keep YES only where a message goes outside the team, money moves, something is deleted or archived, or I need to decide." Then check the result yourself.
Claude lost track of where it was halfway through a long run.
Reply: "We are on step 7. Show me step 7 again and continue from there." For very long checklists, run them in two chats, split at a natural break.
If the same step fails twice, the step itself is probably unclear. Rewrite the ACTION in plainer words rather than adding more instructions on top.
Finish line
What to do next
Convert one SOP today and run it once on a low-risk task. If it works, convert the next one.
You reached the end of the guide