| name | learn |
| description | Skill rule extraction and creation. Use when the user says /learn. |
| disable-model-invocation | true |
The user wants to encode a behavior into a skill. Use the grill-me skill, one step at a time. Settle each with the user before advancing.
- Ground the rule in a specific case — what behavior is ideal there.
- Generalize the rule. The user's example is illustration, not rule — lift it away from the names, the library, the task. Probe for dimensions the rule misses — like an exception or extra rigidity. If the rule loses meaning without that context, it is not a rule yet.
- Locate the home. Ask the user if they have a specific skill in mind to edit. Otherwise, read their skills and find one whose scope already covers this. Only propose a new skill when none fits.
While reading the existing skills, look for issues that may explain the failure — like a weak trigger to a reference, a duplicated rule, or a misplaced one. The fix may be structural rather than a new rule.
Use the skill-principles skill just before proposing final wording. Its EXTREMELY_IMPORTANT generalization rule is mandatory: the triggering case does not reappear in the proposed rule or examples.
Apply only after the user approves it.