My second brain

My Second Brain

Use this skill to help a user operate a personal knowledge vault as a living system, not a dumping ground. The core loop is:

Capture -> Triage -> Compile -> Connect -> Projectize -> Publish -> Review

Defaults

• Target folder: current working directory unless the user provides a path.
• Language: follow the user's language. If unclear, use the vault's existing language.
• Preserve user-written content. Prefer appending update sections or new linked notes over overwriting.
• Do not move application code into the vault. Store knowledge and path pointers only.
• Treat original sources as raw evidence: preserve source links, mark uncertainty, and avoid inventing facts.
• My Take belongs to the user. You may draft suggestions, but clearly label them if they are AI-proposed.

First Read

For an existing vault, read these if present:

1. AGENTS.md
2. HANDOFF.md
3. _System/AGENTS.md
4. _System/Workflows.md
5. _System/Schema.md
6. Dashboard.md

For a folder that is not yet a vault, read references/vault-method.md and references/schema.md, then create the smallest useful structure.

Decision Tree

User wants to process newly collected material

1. Inspect 0 Inbox/, 1 Sources/, and any files the user named.
2. Triage each item:
   • Raw article, transcript, clipping, chat log -> 1 Sources/
   • Reusable concept -> 2 Notes/ with type: note
   • Repeatable procedure -> 2 Notes/ with type: method
   • Active work or experiment -> 3 Projects/ with type: project
   • Public-facing draft -> 5 Outputs/ with type: output
3. Compile valuable sources into concise notes. Keep citations and uncertainty.
4. Add My Take and Open Questions sections where useful.
5. Update related MOCs or create one when a topic has enough notes.
6. Refresh Dashboard.md.
7. Tell the user the 1-3 human decisions or My Takes they should supply.

Read references/workflows.md before doing multi-file processing.

User wants a clear after-collection checklist

Give the minimal executable checklist in the user's language:

1. Triage new 0 Inbox and 1 Sources content.
2. Decide what to archive and what to compile.
3. Turn valuable material into 2 Notes knowledge cards or method cards.
4. Preserve sources and mark uncertainty.
5. Leave My Take and Open Questions.
6. Create or update 3 Projects when knowledge needs action.
7. Create 5 Outputs drafts when knowledge can be published.
8. Refresh Dashboard.md.
9. End with the top 1-3 human decisions.

User wants to initialize a vault

1. Create only the minimal folders: 0 Inbox, 1 Sources, 2 Notes, 3 Projects, 4 Me, 5 Outputs, _System, Attachments.
2. Create or merge _System/Schema.md, _System/Workflows.md, AGENTS.md, HANDOFF.md.
3. Copy or adapt templates from assets/.
4. Add lightweight scripts only if the user wants automation.
5. Do not create large empty taxonomies.

Read references/schema.md and references/vault-method.md.

User wants a vault health check

1. Run local vault scripts if they exist:
   • python scripts/dashboard.py --save
   • python scripts/notes_status.py
   • python scripts/vault_report.py
2. If the vault lacks these scripts, use bundled scripts from this skill with --vault <target>.
3. Summarize:
   • uncompiled sources
   • notes missing My Take
   • unanswered Open Questions
   • draft outputs
   • stuck projects
4. Recommend one next action, not ten.

Read references/dashboard-rules.md.

User wants to save something from the current conversation

This is live capture, not file processing. Keep friction near zero:

1. Distill what the user pointed at (an insight, decision, method, gotcha, or the whole discussion) into a concise note. Do not dump the raw transcript.
2. Route by nature, defaulting to lightweight:
   • Not sure where it belongs -> a seed note in 0 Inbox/ (one file, frontmatter + a few lines). Never block capture on classification.
   • Clear reusable concept/insight -> 2 Notes/ (type: note)
   • Repeatable procedure -> 2 Notes/ (type: method)
   • Belongs to an existing project -> append to that project card's log/decisions instead of a new file.
3. Set source: to something like conversation <YYYY-MM-DD> (<topic or tool>) so provenance survives.
4. The user's own words go in My Take; anything you inferred is labeled AI-proposed. Add 1-2 [[links]] to related notes if obvious.
5. Confirm in one line what was saved and where. Skip dashboard refresh unless asked — quick capture should feel instant.

User wants to turn knowledge into output

1. Gather the relevant notes/projects.
2. Identify the user-owned angle, not just facts.
3. Draft into 5 Outputs/.
4. Mark sections needing user confirmation.
5. Link the output back to source notes and projects.

Bundled Resources

• references/vault-method.md: mental model, principles, and when to split vaults.
• references/schema.md: recommended frontmatter and folder conventions.
• references/workflows.md: capture-to-output SOPs and common user prompts.
• references/dashboard-rules.md: dashboard and participation checks.
• assets/Source.md, assets/Note.md, assets/Method.md, assets/Project.md, assets/Output.md: starter templates.
• scripts/new.py, scripts/dashboard.py, scripts/notes_status.py, scripts/vault_report.py: portable standard-library helpers. Run with --vault <target> if not executing from the vault root.

Activation Limits and the Standing Rule

A skill only fires when the user's request matches its description — it cannot watch conversations in the background and decide on its own that something is worth saving. To close that gap, offer to install a short standing rule in the user's global agent config (~/.claude/CLAUDE.md for Claude Code, ~/.codex/AGENTS.md for Codex), adapted to their language:

## Knowledge Capture
- I keep a personal knowledge vault at <vault path> (skill: my-second-brain).
- When a session produces durable value (an insight, method, decision, or hard-won gotcha),
  proactively ask once: "Want me to save X to your vault?" If yes, use my-second-brain.
- Test: still useful in a month, or in another project? If not, don't ask.
- If I say "save this to my brain" (or equivalent), just do it.

Offer this during vault initialization, and whenever a user wonders why insights aren't being captured automatically. This is advisory (not guaranteed); for a hard guarantee suggest a SessionEnd hook, noting it fires every session and can feel naggy.

Finish Every Session

Before finishing, update HANDOFF.md if present, refresh Dashboard.md when content changed, and report:

• files created/updated
• what was compiled or linked
• what the user needs to decide or write next