A way to hand Claude or ChatGPT a multi-step job (a research brief, a weekly report, a client deliverable) so it checks and fixes its own work, and you only sign off at the one step that's hard to undo. Inspired by a widely shared X post on loops and graphs, rewritten for people who don't code, with our own worked example.
Write the job as 3 to 6 parts. Each part gets its own instructions and its own source material, and can be done in a fresh chat.
Here is a job I do often: Describe the job here. Split it into 3 to 6 parts. For each part give: a short name, what it starts from, what it must hand back, and what it must not touch. Each part should be doable in a new chat with only its own material.
Why fresh chats: parts that share one long chat start copying each other's mistakes.
For each "and then" in your plan, ask: does this part actually read what the last part produced?
- Yes: it has to wait.
- No: it can run at the same time.
In our example, three of four "and then"s were fake. Only the final merge had to wait. This is where most of the time comes back.
A check has to be able to say "no", and you should be able to tell right away if it passed. See "Real checks and fake checks".
For this part, write 2 or 3 checks I can verify by looking, not by trusting. Each check must be able to fail. Example: "every number has a link and a month", not "looks accurate".
When a part fails its check, don't redo everything. Send that one part back with a note (see "Send-back note") and leave the rest alone.
- Three tries at most. If it still fails, the plan or the check is wrong, not the AI. Fix that instead.
- Easy to undo (a draft, a note to yourself): let it through once its checks pass.
- Big but reversible (a document others will read, a file you'll replace): needs every check to pass, and you spot-check one.
- Hard to undo (sending, publishing, paying, deleting): never automatic. This is your approval step.
Decide by how hard it is to undo, not by how sure the AI sounds.
When a send-back fixes something, write the lesson into the split instructions for next time ("every number needs a link and a month"). The same mistake shouldn't come back next week. This is the step most people skip.
| Fake check | Real check |
|---|---|
| "Looks good" | Every claim has a link you can click |
| "The AI says it's confident" | Numbers match the linked page |
| "No errors" | Under 80 words; has the three required sections |
| "I skimmed it" | Only the files listed in the plan were changed |
A fake check can't fail, so it never catches anything.
Copy this when a part fails:
Part: which part Result: did not pass Why: which check failed Evidence: the exact line that fails Scope: change only this; leave everything else as it is
Job: a one-page brief on MarkItDown for a newsletter.
- Split: what it is, how to install on a Mac, what it can't do, license and popularity. None needed another's output, so all four could run at the same time; only the merge had to wait.
- Checks: every fact has a link and a month; numbers match the page; each part under 80 words.
- First round: three passed. "License and popularity" said "about 190k stars" with no link, so it failed.
- Send-back: only that line, with the note above. Passed on the second try.
- Merge: pasted together in a fixed order, no rewriting.
- Approval: publishing can't be undone, so that was the one step we read and approved.
- Rule added: "every number needs a link and a month".
Tried on Claude, 2026-11. We ran the parts one after another in a single chat, each with only its own material; we didn't open separate chats.
put·here is made by one person with AI doing most of the work. Our version: small, reversible changes go through on their own; anything the public sees gets a second AI to review it first; and releases, money and publishing wait for the founder's own word. Rules we settle on get written into a decision log, which is our "turn fixes into rules" step.
Pick one job where you read every AI answer. Split it (see "Step 1 · Split the job into parts that stand alone"), write one real check per part, and mark the single step you'll approve.
Lay out the parts as cards in one put·here canvas, with the checks under each and the approval step at the end. Next time you run the job, the map is already there.
Pick AirDrop to send it to your Mac. No AirDrop? Email it to yourself or save it to Notes.
The phone app can’t import yet. Open this link on a Mac to put it in put·here.