| name | save-learned-lessons |
| description | Capture durable lessons from the current session — user preferences, project conventions, gotchas, and reusable techniques — and route each to its right home (the memory system, project/global CLAUDE.md or AGENTS.md, or an existing/new custom skill). Use when the user wants to save lessons, capture what was learned, or mentions "save learned lessons". |
Save Learned Lessons
Review this session and extract the durable lessons worth keeping — things that
help a future session, not one-off details. Look for:
- Corrections & preferences the user gave (how they want work done, tools, tone).
- Project facts & constraints you discovered that aren't obvious from code or git.
- Gotchas — non-obvious failures, footguns, or environment quirks you resolved.
- Reusable techniques — an approach that worked and would apply again.
Skip anything already recorded, derivable from code/CLAUDE.md/git, or relevant
only to this one conversation.
For each lesson, pick exactly one destination:
| Lesson type | Destination |
|---|
| User preference, feedback, project/reference fact | Your memory system — one file per fact in its documented format, plus a pointer in the memory index |
| Preference that applies to every project | Global ~/.claude/CLAUDE.md |
| Convention specific to this project | The project's CLAUDE.md / AGENTS.md |
| Reusable technique for a domain | The most relevant existing custom skill — update it. Create a new skill only if none fits. |
When a lesson is really a whole repeatable workflow that warrants its own skill — not a single fact — use the extract-a-workflow-into-a-skill capability in your set rather than hand-rolling it here.
Before writing, check for an existing entry that already covers the lesson and
update it instead of duplicating.
Then present every proposed change grouped by destination — path plus the exact
text to add or edit — and wait for approval. On approval, apply the changes and
report what changed.
