| name | capturing-knowledge |
| description | Captures durable learnings into a second-brain knowledge base (a folder with a conventions.md or KB-declaring CLAUDE.md) when conditions warrant - a reusable insight, a framework crystallizing, a constraint discovered, a decision trail completed. Triggers on "save this to my notes/knowledge base", "file that under", "worth keeping", or the natural pause at the end of substantive work where a real learning emerged. Only operates when a knowledge base exists; not for process chatter, debugging logs, or tactical one-offs. |
Capturing Knowledge
Route a learning to its canonical home in the knowledge base and write it. The judgment of whether something IS a learning matters more than the speed of saving it.
When to fire
Take initiative — don't wait for a command:
- A session concluded with a reusable insight — the kind future-you would want to find in six months
- The user signals it: "save this", "remember this", "that's worth keeping", "file that under X"
- A natural pause at the end of substantive work where a real learning surfaced
- A deliverable revealed a reusable lesson distinct from the deliverable itself
Hard suppression — do NOT fire for
- Process chatter, debugging logs, tool mechanics
- Facts already canonical somewhere else (the CRM, the codebase, official docs)
- Anything stale within 30 days
- Restatements of what's already in the knowledge base
- Generic summaries and common knowledge findable anywhere
- Small tactical decisions that won't recur
When in doubt, leave it out. A run that saves nothing is fine.
Step 0: Confirm a knowledge base exists
Before writing anything, locate the knowledge base: a conventions.md, or a CLAUDE.md that declares note conventions, at the working directory or an ancestor. If none exists, do not write a note — you are in an ordinary folder (often a code repo), not a knowledge base. Offer /second-brain:setting-up-knowledge-base and stop. This skill never scatters knowledge notes into arbitrary directories.
The judgment bar
A saved learning must be both:
- Findable — it carries the tags and entity links that future search will use. If you can't predict what query would retrieve it, it isn't crystallized enough.
- Usable — it changes future work. "Fact noted" isn't a learning; "this changes how I'd approach X" is.
Fails either test → don't save; tell the user in one line what you considered and why you skipped it.
Canonical routing
Read the knowledge base's conventions.md for the folder map, then route by what the learning IS:
| Kind of learning | Canonical home |
|---|
| Reusable mental model or decision framework | Its own note under notes/ (or the KB's frameworks location) |
| Tool constraint, API gotcha, integration insight | A reference note for that tool — extend it if it exists |
| Insight about how a person or organization operates | That entity's hub page — add or extend a "How they operate" section |
| The "why" behind a project decision | That project's notes — a decisions note, appended chronologically |
| Research finding spanning sources | The relevant synthesis note |
| Reflection not durable enough for its own file | Today's daily note |
No fit after honest effort → a new note in the most specific existing folder. New folders almost never; state your reasoning when creating one.
Authoring shape
New note or appended section, same form:
## <Title — short, specific>
<2–4 sentences. The insight, stated cleanly. No padding.>
**Context.** What surfaced this — the session, decision, or data point.
**Implication.** What future work does differently because of it.
Apply the KB's frontmatter and link conventions (the organizing-notes skill covers both). Bump modified: when extending an existing note.
Relation to built-in memory
Claude Code's built-in auto memory also responds to "remember this" / "save this," storing terse machine-local facts in MEMORY.md. That's a different store from this knowledge base, and splitting durable knowledge across both fragments it. Rule of thumb: a substantive learning (framework, decision trail, research finding) belongs here, as a real note in its canonical home; a one-line operating preference about how to work with this user can stay in built-in memory. When a capture has real content, prefer the knowledge base and say so.
Report
- Saved: one line — what and where, plus the tags/links applied
- Skipped: one line — what you considered and why it didn't clear the bar
