One of the biggest sources of waste in an AI coding workflow is not that an agent cannot fix a problem. It is that the next agent encounters the same class of problem and starts from zero.
I now use a simple rule: the same mistake is allowed to happen twice.
The first time, solve the task. The second time, solve it again, but also decide where the next agent should have encountered that lesson before it started.
Four places to write a lesson back
1. Session closeout
Capture the conclusion, evidence, risk, and unfinished work. This prevents the next session from treating a confident summary as verified state.
2. Project documentation
Stable facts, current state, and navigation belong to the project. They should not live only inside one conversation.
3. Skills and agent instructions
Repeated execution behavior belongs in a default instruction only when the action is stable. Record what to read first, which boundaries are hard, how to verify the result, and when to stop for human judgment.
4. Periodic review
Instructions age. Periodically remove stale entry points, merge duplicated rules, and fix triggers that are too noisy or too hard to hit.
The boundary matters
Not every incident deserves a new Skill. One-off context stays with the session. Project facts stay with the project. Only behavior that repeats across tasks and has a stable response should become a reusable default.
This is not an agent improving itself. It is a human turning experience into an instruction that the next run can load by default.
A small experiment: take one problem you encountered twice this week. Write a checklist of no more than 20 lines, including one validation method and one stop condition. Then see whether the next agent catches it before repeating the mistake.
Top comments (0)