brain-maintain-knowledge

name	brain-maintain-knowledge
description	Maintain durable knowledge pages using raw sources, organized notes, canonical notes, indexes, logs, backlinks, deduplication, restructure recommendations, and conservative cleanup radar.

Maintain Knowledge Skill

Follow AGENTS.md first. This skill only defines task-specific knowledge maintenance procedures.

Mental model

This repo follows a source-preserving knowledge maintenance pattern:

• Raw capture/source-of-record material lives in 00-inbox/ and is not rewritten during normal processing.
• Organized editable notes and source-derived material live in 10-notes/.
• Compiled durable knowledge is maintained in 20-knowledge/.
• The schema is defined in root instructions and focused skills.
• Human effort should focus on capture, review, and direction.

Default pass

Important: Each operation is independent. A clean result from one operation must not be used to skip another operation.

When invoked without a specific target, run all maintenance operations in a conservative sequence:

1. Read the main navigation and maintenance surfaces:
   • 20-knowledge/index.md
   • recent entries in 20-knowledge/log.md
   • .generated/README.md
   • .generated/unresolved.md, .generated/stale-claims.md, .generated/duplicates.md, and .generated/orphan-pages.md if present
   • .generated/graphify/GRAPH_REPORT.md if present
2. Run Source Check for sensitive, important, stale, or high-impact claims surfaced by indexes, logs, generated reports, or Graphify.
3. Run Lint across the maintained knowledge surfaces.
4. Run Domain Placement Audit across 10-notes/, 20-knowledge/, and active 30-projects/ contexts.
5. Run Navigation Repair, applying small local fixes only when the fix is clear and source-preserving.
6. Run Deduplication, resolving duplicate or overlapping notes while preserving source traceability.
7. Run Restructure Options, returning advisory before/after options in chat.
8. Run Archive Review, returning archive candidates and rationale in chat.
9. Run Cleanup Radar, returning ranked cleanup candidates in chat.
10. Ask the human only when a decision affects broad folder moves, non-trivial duplicate merges, source interpretation, or archival of raw/source material.

Default output:

• Changes made
• Issues found but not changed
• Source checks
• Domain placement findings
• Duplicate findings
• Restructure options
• Archive candidates
• Cleanup radar
• Follow-up decisions needed

Operations

Source Check

• Start from canonical pages, then verify sensitive or important claims against 10-notes/ and, when needed, original 00-inbox/ captures.
• Prefer direct organized/source material over lossy summaries when claims affect decisions.
• Mark unsupported, stale, or contradictory claims for maintenance.
• Do not save new reusable knowledge unless explicitly instructed.

Lint

Look for:

• stale claims
• contradictions
• duplicate pages
• orphan pages
• missing backlinks
• missing source references
• missing provenance fields or mismatches
• unresolved questions
• plain path file references
• mismatches between file path, frontmatter domain, source-folder category, and durable subject

Navigation Repair

• Fix broken or stale links, missing backlinks, index gaps, and plain filesystem path references.
• Convert metadata file references to quoted Obsidian wikilinks.
• Add or repair provenance frontmatter when the correct value is clear from the source context.
• Use generated artifacts such as orphan reports as hints, not canonical truth.
• Update navigation only when it improves findability.

Domain Placement Audit

• Always check whether files live in the right domain.
• Compare path, frontmatter domain, source folders, title, tags, and nearby index ownership.
• Surface concrete before/after move options with affected files.

Restructure Options

• Propose folder or page restructuring when repeated friction, crowding, misplaced notes, overlapping folders or files, or unclear canonical ownership appears.
• Consider move, rename, merge and split recommendations.
• Return concise before/after structure, rationale and affected notes or folders.
• Do not move/rename/merge/split unless explicitly asked.

Archive Review

• Identify stale, superseded, inactive, or one-off material that may belong in _archive/.
• Return archive candidates and rationale in chat instead of creating a generated report.
• Never recommend archiving 00-inbox/ source-of-record material unless the human explicitly asks.
• Prefer "candidate" language unless canonical replacement, source preservation, and link impact are clear.

Cleanup Radar

• Advisory only by default.
• Return cleanup candidates in chat instead of creating a generated report.
• Do not move, delete, archive, rewrite, or generate files unless explicitly asked.
• Use these ranked buckets:
  • Archive candidates
  • Low-signal or low-value content
  • Duplicate or overlapping content
  • Stale generated artifacts
  • Do-not-touch or preserve items
  • Follow-up decisions needed
• Mark uncertainty and weak evidence clearly.
• Prefer a small number of high-confidence recommendations over a broad sweep.

Deduplication

• Treat duplicate or overlapping pages as a normal maintenance concern.
• Detect duplicate titles, overlapping concepts, repeated source summaries, and competing canonical pages.
• Compare overlapping notes before merging.
• Identify the canonical target.
• Merge durable content into the canonical note.
• Preserve source references.
• Add redirects or backlinks from old notes.
• Move deprecated notes to _archive/ only when the canonical target is clear, source preservation is verified, and link impact is understood.
• Update .generated/duplicates.md.
• Update relevant indexes and 20-knowledge/log.md.
• Preserve source traceability and confidence or sensitivity distinctions.
• When overlap is partial, link notes instead of forcing a merge.

Skill-specific judgment

• Keep knowledge pages readable by humans.
• Keep durable pages current-state oriented.
• Use logs and generated reports to surface maintenance needs, not as canonical truth.
• Prefer direct organized/source material over lossy summaries.
• Mark assumptions explicitly.
• Preserve uncertainty and source context when claims conflict.
• Keep the original language and tone of the source material when possible.
• Write metadata file references as quoted Obsidian wikilinks.