kb-learn

/kb-learn

Contents

• 1. Configuration
• 2. Routing
• 2a. Agents
• 3. KB Discovery
• 3a. Concept Evolution
• 4. Update KB
• 4a. Structure Health Check
• 5. User Approval
• 6. Naming and Folders

1. Configuration

On first invocation, check for config at .claude/knowledge-base/config.json.

No config: Tell user "Run /kb-bootstrap to set up your knowledge base." and stop.

Config exists: Read kb_roots. Single entry → use it. Multiple → defer to §4 step 0.

2. Routing

Parse $0 to determine workflow:

• article → reference/article-workflow.md
• topic → reference/topic-workflow.md
• fix → reference/mistake-workflow.md
• compact → Redirect: "Run /kb-compact [path] instead."
• Otherwise → show usage with modes (article/topic/fix), tips, and examples

2a. Agents

Specialized agents isolate high-volume work and enable parallelism. SKILL.md is the orchestrator — spawns agents and chains results. Agents never spawn other agents.

Agent	Model	Purpose	Used by
scouter	haiku	KB discovery via /kb-find (supports --challenge)	all workflows
searcher	sonnet	Web research — search, fetch, structure findings	topic, article
challenger	sonnet	Adversarial web research — finds counter-evidence	topic
assessor	opus	Evaluate claims against KB + web evidence	article, topic

Orchestration Pipeline

1. Spawn scouter(s) + searcher IN PARALLEL
   ├── scouter (normal) → supporting evidence
   ├── scouter (challenge) → counter-evidence
   └── searcher → web findings
2. Spawn challenger (with claims from step 1) → counter-evidence from web
3. Spawn assessor (with all evidence incl. challenger) → verdicts + KB impact
4. Apply changes

Not all steps required every time:

• Article: scouter(s) + searcher → assessor (no challenger)
• Topic: full pipeline including challenger
• Fix: single scouter, no web search

Use agents when output is high-volume, work is parallelizable, or task benefits from model specialization. Use inline when KB is small, single simple claim, or latency matters.

3. Shared Procedure: KB Discovery

All KB discovery goes through the scouter agent. Never use /kb-find directly.

How to spawn scouters

1. Read the kb-find SKILL.md at ${CLAUDE_SKILL_DIR}/../kb-find/SKILL.md
2. Build prompt: scouter body text (after frontmatter) + separator + full kb-find SKILL.md content
3. Spawn Agent with model: haiku, background: true

Why: Skill-internal agents don't get automatic skill injection — the orchestrator must explicitly inject kb-find content.

Spawning patterns

• Bare lookup (topic, fix): single scouter, normal mode
• Claim evaluation (article, deep assessment): two scouters in parallel (normal + challenge)

Per-workflow queries

• article: search by claims and key terms
• topic: search by topic name
• fix: search by described error (with scrutiny)

Empty KB: Skip loading, note KB is empty.

3a. Shared Procedure: Concept Evolution and Legacy Patterns

After loading KB, check for concept evolution — outdated terminology, framing, or versioning. Follow reference/evolution-workflow.md for detection, callouts, old patterns, and legacy notes.

4. Shared Procedure: Update KB

0. KB routing: Multiple KBs → infer target from context; if ambiguous, ask user.

   • Read-only guard: If target KB has [read-only] badge (global KB), do NOT write to it. Instead, create a suggestion file in ~/.claude/knowledge-base/suggestions/ using suggestion.py (see ${CLAUDE_SKILL_DIR}/scripts/suggestion.py --help). Inform the user that a suggestion was created for the global KB. 0a. Skill folder check: If topic has skill/ subfolder, route by content type:
   • "What/where/when" → KB concept notes (lightweight: summary, sources)
   • "How/when to use/procedures" → skill's reference/ or SKILL.md
   • Move overlapping "how" sections from concept notes into skill (no duplication)
   • No skill/ → all content to concept notes

1. Templates from assets/: topic-note, index-note, legacy-note

2. Slim frontmatter: <!-- CANONICAL: Note format (frontmatter, TOC). Other files reference this section, not restate it. -->

   • Topic/index notes: only name + description in frontmatter. No title, summary, confidence, tags.
   • Inline <conf:high>, <conf:medium>, <conf:low> markers on claims are allowed and preserved during edits. Claims without a confidence tag should be treated as <conf:medium> (unverified but not flagged).
   • Body: # {Title} → content sections. No summary paragraph duplicating description.
   • TOC: Notes >100 lines need ## Contents in first 10 lines of body.
   • Legacy notes: name (append " (legacy)") + description (mention replacement).

3. Keep heading accurate after modifications.

4. Cross-references: markdown links with relative paths. Every concept note should link to at least one other note/index. Cross-KB: use @kb-name/path format. Global KBs: use soft references only — see: @namespace.kb/topic (no hard markdown links to read-only KBs).

5. Index updates: Link new notes from relevant index. Create index if 3+ notes and none exists. Use table when section has 3+ items (see reference/folder-structure.md).

5a. Sibling unification: Skip if §4a already ran. Otherwise, if creating a new folder, check siblings per §4a.

6. MANDATORY changelog: Every change to KB files → prepend entry to that KB's CHANGELOG.md. Follow reference/changelog-workflow.md. Do NOT skip.

7. Large notes: If note reaches ~500 lines, suggest /kb-compact <path>.

4a. Shared Procedure: Structure Health Check

Run after any KB update. Reuse KB root listing from step 5 if available; otherwise one ls.

Detection: Group KB root folders by first hyphen-delimited segment. 3+ folders sharing prefix (no existing parent) = candidate. No candidate = stop.

If found: Ask user — "Restructure now", "Skip", "Not applicable".

If approved: Follow reference/folder-structure.md. Run validate_kb.py --quiet to confirm zero errors.

5. Shared Procedure: User Approval

Before non-trivial KB changes, present summary: notes to create, modify, index updates, changelog entry. Group claims by confidence level. Show <conf:low> claims in a separate "⚠ Low confidence" section. Use the AskUserQuestion tool (not plain text) with options: "Apply high/medium confidence", "Apply all (including low confidence)", "Apply selectively", "Cancel".

Exception: Minor changes (adding source URL, fixing typo) may proceed without approval but must be mentioned.

6. Note Naming and Folder Conventions

• Lowercase, hyphen-separated. Topic-prefixed when ambiguous (crdt-operation-based.md, not operations.md). 2-4 words.
• Index: one index.md per folder. See reference/folder-structure.md.

Skill Folders (KB-backed skills)

Topics with "how" content have skill/ inside their concept folder. Symlink from .claude/skills/ makes it discoverable. All skills must be KB-backed. When skill/ exists, concept notes are lightweight (provenance, summary, pointers to skill).

Forward-Only

Conventions apply to new notes by default. Existing notes not moved unless requested. When editing, may normalize frontmatter and update indexes.

Moving files

When relocation requested: mv files → run validate_kb.py → fix reported broken links. Don't rewrite content just to relocate.

7. Validation

After completing KB changes, run the validator to catch broken links, frontmatter issues, and structural errors:

uv run ${CLAUDE_SKILL_DIR}/scripts/validate_kb.py --quiet --json

Fix any reported errors before finishing.