Search Skills

story: e09s01

story: e21s01

Search Skills

HARD GATE — Search results must be ranked by relevance. Do NOT return all matches without prioritization. Use skill metadata (phase, purpose, frequency) to rank.

HARD GATE — Do NOT use external embedding APIs or AI-based semantic search. This is a lexical-only index (ADR: zero external dependency).

Lexical search only — no embedding service (ADR: zero external dependency). The index is a flat markdown file (specs/SKILL-SEARCH-INDEX_LATEST.md) built from every SKILL.md's YAML frontmatter — name, description, and key phrases. No vector DB, no API calls, no network dependency.

When to use

• You're unsure which skill to invoke for a user's request
• At the start of research-first to find pre-existing skills that might solve the problem
• When a user asks "is there a skill for X?"
• Before calling a skill by name, to confirm it's the right one

Pre-flight

• Does specs/SKILL-SEARCH-INDEX_LATEST.md exist? If not, run bash scripts/build-skill-index.sh.
• Is the index fresh? Check its timestamp — if > 24 hours old or after any SKILL.md change, regenerate.

Process

1. Refresh index if stale — Run bash scripts/build-skill-index.sh if specs/SKILL-SEARCH-INDEX_LATEST.md doesn't exist or was modified before the last SKILL.md change.

2. Search the index — Use ripgrep on the lexical index:

   Copy code
   rg -i "<keywords>" specs/SKILL-SEARCH-INDEX_LATEST.md
   

   The index contains each skill's name, description, phase, and key use-case phrases, so natural language queries work well even without embeddings.

3. Rank results — Read the top 3 matches. Evaluate by:

   • Exactness — Does the description literally match the user's intent?
   • Phase fit — Is the skill designed for the current lifecycle phase?
   • Trigger phrases — Does the skill's "Use when" section match the situation?

4. Recommend one skill — Select the single best-matching skill. Provide:

   • The skill name
   • Why it's the best match (citing the description or trigger phrase)
   • What it produces (artifact, dialogue, or state change)

5. Invoke — Call the skill directly or through the orchestrator. If no match found, suggest the closest phase-appropriate skill or using-bigpowers as a general entry point.

Index Format

specs/SKILL-SEARCH-INDEX_LATEST.md contains one section per skill:

markdownCopy code
## <skill-name>
- **Description:** <from frontmatter>
- **Phase:** <lifecycle phase>
- **Triggers:** <key phrases from description>
- **Keywords:** <extracted terms>

Why Not Semantic Search?

• Zero network dependency — works fully offline
• Zero cost — no API keys, no usage limits
• Instant — ripgrep on a local file is sub-second
• Deterministic — same query always returns same results
• Auditable — you can read the full index

Verify

→ verify: test -f specs/SKILL-SEARCH-INDEX_LATEST.md || (bash scripts/build-skill-index.sh && test -f specs/SKILL-SEARCH-INDEX_LATEST.md)