Skill: Discover Skills

Purpose

Help the Agent identify missing skills for a task and recommend concrete installation steps. This skill discovers candidates and suggests what to install, but does not install or inject skills automatically.

---

Core Objective

Primary Goal: Provide the Agent with top 1-3 skill recommendations and exact installation commands to fill capability gaps.

Success Criteria (ALL must be met):

1. ✅ Discovery performed: Searched local skills/INDEX.md and manifest.json, or external catalogs if requested
2. ✅ Best matches identified: Selected 1-3 skills that match task requirements by name, description, and tags
3. ✅ Recommendations explained: Provided rationale for why each skill matches the current task
4. ✅ Install commands provided: Included exact installation command for each recommended skill
5. ✅ No auto-install: Did not execute installation commands without explicit user confirmation

Acceptance Test: Can the Agent or user install the recommended skills using the provided commands without additional research?

---

Scope Boundaries

This skill handles:

• Discovering skills from local indexes or external catalogs
• Matching task requirements to skill capabilities
• Recommending top 1-3 skill matches with rationale
• Providing exact installation commands

This skill does NOT handle:

• Installing skills automatically (use install-rules or manual installation)
• Curating or auditing existing skills (use curate-skills)
• Refining or designing skills (use refine-skill-design)
• Injecting skill content into Agent context (handled by Agent runtime)

Handoff point: When recommendations are provided with install commands, hand off to user for installation decision or to install-rules for automated installation (with user confirmation).

---

Use Cases

• Initial bootstrap: The Agent starts with only this skill, then recommends which skills to install for the current task.
• On-demand extension: When a task requires a Skill that is not available, suggest the best matches and how to install them.
• Capability discovery: Help users find relevant skills from local indexes or public catalogs.

Behavior

1. Discovery: Prefer local skills/INDEX.md and manifest.json for the capability map. If the user asks for external options and network access is available, search a public catalog (e.g. SkillsMP).
2. Matching: Match the user task to name, description, and tags; select the best 1–3 skills.
3. Recommendation: Explain why each skill matches and whether it is local or external.
4. Installation guidance: Provide the exact install command for each recommended skill (e.g. npx skills add owner/repo --skill name).
5. Confirmation: If the user asks the Agent to install, request explicit confirmation before running any command.

Input & Output

• Input:
  • Description of the current task.
  • Optional: list of already installed skills.
  • Optional: allowed sources (local only vs public catalogs).
• Output:
  • Recommended skills with rationale.
  • Install commands for each recommendation.

Restrictions

Hard Boundaries

• No auto-install: Do not execute install commands without explicit user confirmation.
• No auto-injection: Do not fetch or inject remote SKILL.md content automatically.
• No bulk discovery: Avoid listing large catalogs; return only the top 1–3 matches.

Skill Boundaries (Avoid Overlap)

Do NOT do these (other skills handle them):

• Installing skills: Executing installation commands or modifying skill directories → Use install-rules or manual installation with user confirmation
• Curating skills: Auditing skill quality, detecting overlaps, or scoring skills → Use curate-skills
• Refining skills: Designing, restructuring, or improving existing skill content → Use refine-skill-design
• Injecting content: Loading skill content into Agent context or runtime → Handled by Agent runtime system

When to stop and hand off:

• User says "install it" or "add that skill" → Provide install command, request confirmation, hand off to install-rules or manual installation
• User asks "how do I improve this skill?" → Hand off to refine-skill-design
• User asks "are my skills good quality?" → Hand off to curate-skills
• Recommendations provided with install commands → Discovery complete, await user decision

Self-Check

Core Success Criteria (ALL must be met)

• Discovery performed: Searched local skills/INDEX.md and manifest.json, or external catalogs if requested
• Best matches identified: Selected 1-3 skills that match task requirements by name, description, and tags
• Recommendations explained: Provided rationale for why each skill matches the current task
• Install commands provided: Included exact installation command for each recommended skill
• No auto-install: Did not execute installation commands without explicit user confirmation

Process Quality Checks

• Relevance: Are the recommendations strongly related to the current task?
• Actionable: Are install commands concrete and correct?
• Consent: Did the Agent avoid running installs without explicit confirmation?
• Source priority: Did I prefer local indexes before searching external catalogs?
• Conciseness: Did I limit recommendations to top 1-3 matches, avoiding bulk catalog listings?

Acceptance Test

Can the Agent or user install the recommended skills using the provided commands without additional research?

If NO: Recommendations are incomplete. Verify install commands are correct and include all necessary parameters.

If YES: Discovery is complete. Await user decision on installation.

Examples

Example 1: Recommend a local skill

• Scenario: User asks for a standardized README.
• Steps:
  1. Agent checks skills/INDEX.md.
  2. Finds generate-standard-readme.
  3. Recommends it and provides: npx skills add nesnilnehc/ai-cortex --skill generate-standard-readme.

Example 2: Edge case - no local match

• Scenario: The task is niche and no local skill matches.
• Expected: Offer 1–3 external suggestions (if allowed) with install commands, or say "no match found" and ask for clarification. Do not install automatically.

---

Appendix: Output contract

When this skill produces recommendations, it follows this contract:

Element	Requirement
Count	Top 1–3 matches only; no bulk catalog listing.
Per skill	Name, rationale (why it matches), install command (e.g. npx skills add owner/repo --skill name).
Source	Prefer local skills/INDEX.md and manifest.json; external only when user asks and network available.
Interaction	Do not run install commands without explicit user confirmation.