capture

Lore Capture

Capture domain knowledge into the lore knowledge base.

Process

1. Identify the knowledge. Either:

   • Use $ARGUMENTS if the user provided a knowledge statement
   • Extract from recent conversation context if invoked without arguments

2. Classify the knowledge. Determine where it belongs:

   • Lore entry: Domain facts, business rules, entity definitions, decisions with rationale, domain vocabulary — anything declarative about how the domain works
   • CLAUDE.md rule: Imperative instructions — "always do X", "never do Y", coding conventions
   • Skill instruction: Workflow logic, process steps, automation recipes

3. If it's a lore entry, determine:

   • Title: Clear, descriptive name
   • Type: concept, entity, rule, role, decision, or glossary
   • Domain: Which subdomain it belongs to (if any)
   • Confidence: verified if directly stated by the user, inferred if derived from code or context
   • Source: Where this knowledge came from (e.g., "human:brainstorming", "codebase:path/to/file")

4. Check for existing entries. Use lore_search to see if this knowledge already exists. If it does:

   • Show the existing entry
   • Ask if this is an update/correction or a separate entry
   • If updating: use lore_write to overwrite with the new content

5. Generate search keys. Before calling lore_write, generate 5-10 search keys for the entry: synonyms for key terms, alternative phrasings of the main concept, and 2-3 questions this entry would answer. Pass these as the search_keys parameter.

6. Write the entry. Use lore_write with all required fields including search_keys. Present the written entry to the user for confirmation.

7. If it's NOT a lore entry, explain where it should go instead:

   • CLAUDE.md rules → "This is an imperative instruction. Add it to CLAUDE.md or .claude/rules/"
   • Skill instruction → "This is workflow logic. Create or update a skill definition."

Important Rules

• Every entry must trace to a source — never invent knowledge
• When the user directly states something, confidence is verified
• When derived from code analysis, confidence is inferred
• If uncertain about the type, ask the user