Fineuralab
Darwin Skill Improvement Loop
A practical loop for improving AI Skills with task sets, failure logs, revisions, and regression checks.
AI Skills guide
Darwin Skill Improvement Loop
Darwin-style work treats a skill as something that can be evaluated. The key is to compare versions with real tasks instead of editing instructions by instinct.
Create a task set
Collect a small set of representative tasks before editing the skill.
- One easy case.
- Two realistic daily cases.
- One edge case.
- One case the current skill handles poorly.
Record failures
A useful failure log names what went wrong, why it matters, and which instruction might fix it.
- Missing step.
- Unsafe assumption.
- Overbroad trigger.
- Output format drift.
Keep or roll back
A revision should survive the old tasks and improve the target failure. Otherwise it is just movement.
- Run before/after examples.
- Keep changes small.
- Do not optimize for only one sample.
- Write down why the change was kept.