Loop/Prompt Library / Loops from builders on X
🔁 A big AI teaches a small one
A strong AI writes and improves instructions for a smaller, cheaper AI until the small one does a repeat job well.
Use it when: You have a simple job you repeat a lot and want a cheaper or offline model to handle it.
🛑 Stops when: The small model passes your test examples, or the round limit is hit.
Copy this loop
Job: [repeat task, e.g. 'sort support emails into 5 types']. Here are [10] example inputs with the right answers: [examples]. You (the strong AI) write instructions for the small model [name]. Run the small model on the examples and score it. Look at what it got wrong and improve the instructions. Keep [3] extra examples hidden from yourself until the end, then test on them. Stop when the small model gets at least [9] of 10 right and does well on the hidden ones, or after [6] rounds. If it never gets there, tell me this job may need a bigger model.
Review every change the agent makes before you keep it. Set a budget or round limit — the prompt includes one.
Found on X · by @easyldur. Prompt rewritten in plain English by FixKit AI.
More loops like this
One useful thing every night
You give the AI a short mission, and each night it picks one task that moves you closer and leaves it for you to check in the morning.
Two watchers keep your agent healthy
Two separate AI helpers read your main agent's logs, find what broke, and must both agree on a fix before it is used.
Weekly skill cleanup
Once a week, the AI grades every saved skill or saved prompt it uses, removes the dead ones, and merges the repeats.
Overnight drafts that learn from results
Each night the AI drafts tomorrow's post from your content plan, and reads your notes about what worked before it writes the next one.
Split the work, keep only what passes
Splits a big job across several AI helpers working side by side, checks each one's work, and keeps only the work that passes.
Practice a website task until it's reliable
The AI tries a website chore, learns from each failed try, and keeps adjusting until it has steps that work every time.