Compounding loop
After each piece of AI work, writing down what you learned where the AI will read it next time, such as a line in its instructions, a saved example or a new check, so you don't catch the same mistake twice. Named "compound engineering" by Kieran Klaassen at Every.
Also called: compound engineering · compound knowledge · compound writing · lessons learned file · write it down for next time
Connections
Understand first
Ideas this one builds on.
Related
Often used together or easily confused.
Resources that teach it
Competency
Find compounding loop on the commons map ↗
You may know it from
- military and project management an after-action review After each operation the team asks what was supposed to happen, what did happen and what to do differently, and the answers change the next plan.
- cooking a kitchen's prep notes A cook who burns the sauce once writes 'lower heat after the cream goes in' on the recipe card, so nobody burns it again.
In products
The same idea, under different names in different tools.
| Provider | What they call it |
|---|---|
| Every | Compound Engineering plugin; Compound Knowledge plugin |
In plain words
Most people fix an AI’s mistake in the chat and move on. Next week the mistake comes back, because nothing the AI reads has changed. A compounding loop adds one step at the end of each job: decide what you learned, and put it somewhere the AI will see next time.
That place might be your custom instructions, a skill’s checklist, a saved good example, or a short “things to avoid” note in a project. A useful test, from Every’s compound engineering plugin: would someone repeat the mistake if this note disappeared? If not, it isn’t worth saving.
Kieran Klaassen at Every named the idea for software work in 2025. Austin Tedesco, also at Every, applied it to office work, and Katie Parrott to writing.
Example
A grants officer asks AI to draft funder reports. Twice it calls a programme “free” when it is low-cost. She adds one line to her project instructions: “Our programmes are low-cost, never free.” The error stops, and the next report needs one fewer edit.
How it connects
- Builds on custom instructions and skills, the places where lessons get saved.
- A saved check you run every time is a small eval.
- Differs from memory: you choose what gets written down and where.
Try it
After your next AI task, write one sentence about what you corrected. Add it to your instructions or project, then run the same kind of task again and see whether the correction held.
Read this card as plain Markdown: /concepts/con-compound-loop.md. Know a parallel from your own field, or a product that implements this? Add it.