Recurring AI Tasks: Turn Your Best Prompt Into a Loop
An AI loop is a recurring prompt with the parts that let it improve: an objective, acceptance criteria, a check that tests the output against those criteria, a stop rule, a named person who signs off, and a feedback step that changes the instructions whenever you correct the output by hand. To turn a prompt you rerun every week into a loop, write those parts on one page beside the prompt, run it, and fold every hand correction back into the instructions. For recurring AI tasks, that’s the difference between a draft you fix every Friday and one that needs less fixing each week.
A Prompt You Rerun Every Week Is Already Half a Loop
Picture the Friday status digest. You paste in the RAID log export and your Thursday notes, the AI tool drafts the update, and you spend twenty minutes fixing it. Dates in the wrong format. The one red risk buried in paragraph three. A tone that’s cheerier than the project.
Next Friday you make the same three fixes. The corrections live in your head, so the tool never learns them and you never stop making them.
Boris Cherny, who created Anthropic’s Claude Code, told a Sequoia audience in May that he has dozens of loops running: repeat jobs that keep his code changes passing their tests, fix flaky tests, and gather feedback from Twitter into themes every 30 minutes. “I sort of feel like loops are the future at this point,” he said. The same idea works for a status digest, a risk review or a meeting summary. You already have the prompt and the inputs. The missing half is the check and the feedback.
Loops Can Travel Between AI Tools
On 22 October Anthropic handed Agent Skills, the format it created for packaging instructions an agent can reuse, to the Agentic AI Foundation under the Linux Foundation. Anthropic says more than 70 agent platforms have adopted the standard since December 2025, including ChatGPT, GitHub Copilot and Cursor. A skill is a folder with one instruction file, plus any reference material the task needs.
That matters for loops because a loop written down once can become a saved skill, a project instruction or a page in a shared doc, depending on what your organisation’s tools allow. If they allow none of those, it still works as a document you paste from.
The Loop Canvas: Six Zones on One Page
Write the loop on a single page with six zones. Each one answers a question the prompt alone leaves open.
Objective and acceptance criteria: what the output is for, and the tests a good one passes. The AI Agent Charter Check helps you find the gaps in them.
Inputs: the files and notes the tool works from, and how fresh they have to be.
Instructions: the prompt itself. A prompt you already trust, from your own library or a pack such as the Status Reporter, goes straight in here.
Check: a second prompt that tests the draft against each criterion. Run it in a fresh chat, so the draft is marked by a different conversation from the one that wrote it. If you can’t write a check for a criterion, the criterion is too vague.
Stop rule and sign-off: how many redrafts the tool gets, when it stops and asks you instead, and who sends the output.
Feedback: how your hand edits change the criteria and instructions before the next run.
A Worked Loop Canvas for a Recurring AI Task: The Friday Digest
Here’s the Friday digest as a canvas (illustrative):
Objective: a digest the sponsor can read in two minutes and act on.
Acceptance criteria: opens with any red item; every date written as day and month; no item without an owner; risks that moved this week come first; under 250 words.
Inputs: this week’s RAID log export, the milestone tracker, my Thursday notes. Tracker no more than a day old.
Instructions: draft the digest from the inputs in the order the criteria set, in plain language, with no commentary on the team.
Check: in a fresh chat, the tool marks each acceptance criterion pass or fail against the draft. I then check two facts against the tracker.
Stop rule: if any criterion fails, the drafting chat gets the failures and redrafts once. If it still fails, it hands me the draft with the failures listed. If a red item has no owner or the tracker is older than a day, it stops and asks me rather than drafting.
Sign-off: I send it. The tool never sends direct.
Feedback: every edit I make by hand becomes a new criterion or instruction line before next Friday.
The check and stop rule make it a loop inside each run as well as across the weeks: tested, sent back once, then handed to you. Claude Code, Anthropic’s coding tool, has a command built the same way: you set a finishing condition, a separate model checks it after each turn, and you can cap the turns.
Now fill the same six zones for the prompt you rerun most often. Start with acceptance criteria: write down the three fixes you make every time, and you have your first criteria. Before you run it on live material, check your organisation’s policy on pasting client or commercially sensitive project data into a third-party AI tool.
Run It Four Times Before You Trust It
The feedback zone does most of the work. After four runs, your hand edits should be smaller. If they aren’t, look at the criteria first: a fix you keep making that has no criterion is the gap. How to Test a New AI Model on Your Own Project Work applies the same idea to choosing a model.
Keep the sign-off zone even when the loop is running well. A check catches format and order. It won’t catch the steering committee mood shift you heard about in the corridor, and the sign-off is what keeps your name on what goes out.
Once a loop holds steady for a month, it’s worth saving as a skill or shared instruction so a colleague can run it too. AI Skills for Project Managers: Moving Beyond Prompts covers that step up from prompt to skill.
Frequently Asked Questions
What is an AI loop? A recurring prompt with an objective, acceptance criteria, a check, a stop rule, a sign-off and a feedback step that updates the instructions each time you correct the output by hand.
How is a loop different from a saved prompt? A saved prompt repeats the same instructions; a loop tests its output against written criteria and changes its instructions based on what you had to fix.
Can I use Agent Skills in Microsoft Copilot or Gemini at work? It depends on what your organisation has enabled. The loop works as a plain document either way, so write it first and decide where to keep it later.
You already have the prompt. Add the check and the feedback, and the Friday fix gets shorter every week.
Which fix do you make to your AI drafts every single week?
Yes, AI helped me to write this :)