Everyday productivity
Loops that check AI work for you
You will have a loop design with a goal, a test, a limit and a stop rule, tried on one real task.
Before you start
- A task where you can tell good from bad
- An assistant that can follow a multi-step request
The idea
A loop is draft, test, improve, repeat. The test is the important part. Without a clear test, a loop just produces more text.
Step 1: Write a goal and a test
- Goal: what should the result do?
- Test: three yes-or-no questions you can check, such as “Is it under 100 words?”
- Limit: the maximum number of rounds, for example three.
Step 2: Run the loop
Prompt you can copy
Task: [task]. Write a draft. Then check it against these tests: [tests]. If any fail, improve the draft and check again. Stop after three rounds or when all tests pass. After each round, tell me which tests passed.
Step 3: Check the result
The assistant checking its own work is helpful but not final. Test the result yourself, especially the facts.
Set the limits
- Limit the number of rounds.
- Do not let a loop send, post or spend anything.
- Stop and ask if it is stuck on the same test twice.
Common mistakes to avoid
- Looping without a clear test
- Letting it run without a limit
- Letting the loop act in the real world
- Trusting the loop’s own pass mark
Final checklist
- My loop has a goal and a test
- It has a round limit
- I checked the final result myself