| name | learning-capture |
| description | Decide whether a task produced reusable knowledge worth recording — repo memory, session memory, skill candidate, failed approach, prediction error. Use after non-trivial work concludes. |
Learning Capture
Most tasks teach nothing reusable. A few teach something worth keeping. Recognizing which is which keeps the memory useful instead of bloated.
When to use
- A non-trivial task just concluded — success or failure.
- An approach worked or failed in a way that surprises you.
- A user correction or confirmation revealed a preference, constraint, or convention.
Procedure
- Ask the surprise question. Was anything about the outcome non-obvious from the inputs? If no, there is probably nothing to record.
- Pick the right memory tier. Per-task ephemeral → no record. Per-session preference → session memory. Cross-session pattern → repo or user memory. Reusable cognitive scaffold → skill candidate.
- Capture prediction error. If you predicted X and observed Y, the delta is the lesson. Record both predicted and actual; the delta is what teaches.
- Record the why, not just the what. A rule without a reason ages badly; with a reason, future-you can judge whether it still applies.
- Refuse to record what is already in the code. If reading the code answers the question, the code is the memory.
Output
A short capture: what was learned · why it matters · where it applies · where it lives. Nothing if the task taught nothing.
Anti-patterns to avoid
- Recording every task as if it were significant — bloat hides the real lessons.
- Saving facts that the next reader of the code will derive in five seconds.
- Capturing the corrected behavior without the reason — the rule decays without context.
