using-second-brain

<brain_architecture> The Second Brain is stored in docs/second-brain/ (or your chosen Obsidian vault). It has three distinct layers (Karpathy pattern):

1. raw_sources/: IMMUTABLE. User-provided articles, logs, PDFs, meeting notes. You MUST READ from here but NEVER modify these files.
2. artifacts/: IMMUTABLE (write-once). LLM-generated plans (artifacts/plans/) and specs (artifacts/specs/). Created by writing-plans and brainstorming skills. Once written, NEVER modified.
3. wiki/: LIVE SYNTHESIS. A structured, interlinked directory of markdown files. You own this layer entirely. You create, update, and maintain cross-references here. This is where knowledge compounds.

Schema (Configuration): Format rules live in wiki/_SCHEMA.md. Behavioral rules live in .sumela/SKILL_REGISTRY.md and this skill file.

Reliability invariant: For this skill, Second Brain correctness and recoverability are more important than token reduction. Do not remove routing tiers, write gates, parity checks, or memory-sync steps solely to reduce context cost; simplify only when behavior remains at least as safe. </brain_architecture>

<core_indexes> The wiki/ directory has FIVE special files (frontmatter-free, governed by _SCHEMA.md):

• _INDEX.md: Human-optimized content catalog. Read this FIRST every session for navigation.
• _SEARCH_INDEX.md: Agent-optimized search index. A table with per-page Type, Tags, Key Terms, and Summary columns. Use this for targeted search: match query terms against Key Terms/Tags → read only matched pages. Scales to hundreds of pages with zero runtime dependencies. Updated during every ingest/code-commit/lint.
• _LOG.md: Chronological append-only log. Format: ## [YYYY-MM-DD] type | topic (parse-friendly). Type whitelist: ingest | query | lint | code-commit | decision | migration. NEVER rewrite past entries — they are immutable history.
• _LOG.md note: _SCHEMA.md is the canonical whitelist source; evolve is a valid log type for self-improvement applications.
• _SCHEMA.md: Canonical format source. Read this BEFORE writing any wiki page. Defines page types, frontmatter schemas, naming, templates, and cross-ref conventions.
• _improvement-queue/: Self-improvement queue — a directory, one IMP-*.md per signal. Managed by self-improvement-curator skill and reviewed via /evolve. At session start, surface the pending count by globbing IMP-*.md (never scan the whole dir — the README.md example would inflate the count). Do NOT modify as part of normal wiki operations — it has its own workflow. </core_indexes>

<trigger_conditions> Determine which workflow to execute based on these signals (highest specificity wins):

• Session start → bootstrap reads + parity + queue count are governed by .sumela/sumela-prompt.md (<session_bootstrap>). This skill is invoked LAZILY when an information gap appears or one of the triggers below fires.

• User says "save", "add to wiki", "ingest", "note this", "remember this" (or the equivalent in any language) → INGEST workflow.

• User drops multiple sources or says "batch ingest", "process all" (or the equivalent in any language) → BATCH INGEST workflow.

• User says "clean the wiki", "lint", "health check" (or the equivalent in any language) → LINT workflow.

• Ad-hoc decision during conversation — User makes or confirms an architectural/technology decision outside of a coding task (e.g., "let's use Redis", "select this pattern") → DECISION CAPTURE workflow.

• User says "save this", "save this source", "ingest that URL" (or the equivalent in any language), or a lint data-gap suggestion is approved → WEB SOURCE CAPTURE workflow.

• Threshold trigger (ingest-based): After writing a new ingest entry to _LOG.md, count total ingest entries. If count is a multiple of 5 (5, 10, 15, ...), ASK the user: "After the last 5 ingests, I recommend a full lint — should I run it?" — NEVER auto-run lint.

• Threshold trigger (time-based): At session start, after reading _LOG.md, check the date of the last lint entry. If 7+ days have passed since the last lint, ASK the user: "It's been 7+ days since the last lint — I recommend a health check. Should I run it?" — NEVER auto-run lint.

• Ingest with image references → the standard INGEST workflow activates IMAGE READING WORKFLOW (operation 8) as a sub-step when raw_sources/assets/ contains files referenced by the new source. </trigger_conditions>