Task-to-skill matcher

Purpose: Provide one consistent flow for choosing the best installed skill for a user task. Use this when the user asks "what skill should I use?", asks to "pick the best", or when skill choice is unclear.

When to use this skill

• User asks for skill recommendation or best-skill selection.
• Task-to-skill match is uncertain.
• User asks to auto-pick skills for a multi-step or batch run.

Do not use this skill

• If user names a specific skill to run, load that skill directly and skip this matcher.

Sources and scope

• Read installed-skill candidates from the current platform's context loader/index. Use the platform registry documentation when available.
• Cursor: the workspace Cursor context index or skills section.
• Claude Code: the workspace Claude context file.
• Codex: the workspace AGENTS.md.
• Other platforms: the workspace context source identified by the platform registry.
• For complementary public-skill suggestions, check whether ls-find-public-skills is available to the current client. If available, load it and delegate the public discovery flow. If unavailable, report "Public-skill discovery is unavailable. Recommendations: none." and continue installed-skill selection; do not read the public index as a fallback or install a missing package automatically.
• For framework-generated catalogs, prefer ls/docs/_generated/skill-taxonomy.json and ls/docs/SKILLS.md; both are sorted by sort_priority then skill ID and expose class, tags, and pack membership.

Rule ownership

This skill owns task-to-skill selection behavior. TASK_SKILL_MATCHING.md and SKILLS_AND_RULES.md are public references, while this skill carries the runtime matching flow.

• Do not add always-loaded platform rules for every skill; keep platform context compact and route users to generated catalogs or this matcher.
• If a user names a skill directly, load that skill and skip matching.
• For batch work, choose a skill sequence once at the beginning unless the user asks for parcel-by-parcel prompts.

Mode detection

• Treat as batch when user request includes multiple distinct subtasks, or says "batch", "multiple steps", or "run the whole thing".
• Otherwise treat as single task.

Workflow

1. Collect intent

   • Extract user intent and any task constraints.
   • If user already named a skill, stop matching and load that skill.

2. Rank installed skills

   • Compare intent against each installed skill's "when to use" text.
   • Rank by relevance (keyword and description relevance).
   • If uncertain, prepare top 3 candidates with one-line "why it fits".

3. Single-task flow

   • If one clear installed match exists, ask once: "Use this skill?"
   • Show the installed match and "Use this skill?" first, then request complementary public suggestions only if ls-find-public-skills is available, using the availability rule above.
   • Same turn: include up to 3 recommendations returned by discovery, preserving its index-status disclosures and any pending user question. If discovery returns no recommendations, report that result; it does not prevent installed-skill selection.

4. Batch / long-running flow

   • Prompt once at start with options:
     • Auto-pick best skill for the whole run.
     • Parcel-by-parcel prompts.
     • Parcel auto-pick.
   • If auto-pick is chosen, show planned skill sequence first (best-effort), then proceed without repeated skill prompts.
   • If user chooses parcels and phases are unclear, propose one parcel (whole task), then ask prompt-vs-auto-pick for that parcel.

5. No installed fit

   • State that no installed skill is a good fit.
   • If ls-find-public-skills is available, offer up to 3 returned complementary public skills to import; otherwise report the unavailable/no-recommendations result above.
   • Optionally suggest creating a new skill with ls-create-agent-skills.

Public index rules

• Do not implement a separate stale-index policy in this matcher. ls-find-public-skills owns the public index, the standard last-refresh reminder, stale detection, user prompt, refresh, and scrub behavior.
• Preserve discovery's returned last-refresh status (date and age when available), availability and freshness disclosures, and any pending user question. Do not calculate a separate freshness decision, reconstruct its prompts, or perform its maintenance sequence in this matcher.
• For complementary suggestions: return up to 3 public skills from the discovery result, each with one-line fit reason.
• To import suggestions, point user to ls-import-external-skills (or run it if user asks).

Output style

• Keep outputs short, actionable, and consistently structured.
• For uncertain match: show top 3 installed candidates, each with one-line reason, then ask user to choose or say "pick the best".
• For single clear match: ask once, then show complementary public options in the same response.
• For public complementary suggestions, use available enriched fields from the public index entries returned by ls-find-public-skills:
  • Prefer summary_short over raw description.
  • Include notable requirements or risk_flags in one short line.
• Rendering fallback:
  • Rich/basic markdown: numbered list with labeled fields.
  • Plain text/ascii: numbered list with Skill:, Why:, Risks:.

References

• TASK_SKILL_MATCHING.md - Framework task-skill matching guidance.
• SKILLS_AND_RULES.md - Framework skills and rules documentation.
• PLATFORM_REGISTRY.md - Framework platform registry documentation.
• SKILL_DISCOVERY.md - Public skill index and discovery workflow managed by ls-find-public-skills.
• ls-find-public-skills - Complementary public-skill recommendations.
• ls-import-external-skills - Import selected public or local skills.
• ls-create-agent-skills - Create a new skill when no installed or public skill fits.