Skill Discovery

Skill: skill-discovery

classification: Core Universal
tier: T0 (System Behavior)
confidence: 10/10
source: system-mandatory
dependencies: pre-action, memory-keeper aliases: skill-discovery, automatic-skill-discovery

Purpose

Skill Discovery ensures the agent has the correct domain expertise for every task. It performs two critical functions:

1. Internal Auto-loading (Phase 0): Automatically identifies and loads installed skills based on task context.
2. External Suggestion: Proactively identifies gaps and suggests relevant community skills from skills.sh.

Phase 0: Automatic Classification

Execute BEFORE any tool call.

Algorithm

1. PARSE request to identify task type and domain.
2. CLASSIFY by task type (not language):
   • Implementation — Writing code in any language
   • Testing — Writing tests, test fixtures, test harnesses
   • Writing/Documentation — Prose, READMEs, ADRs, runbooks, API docs
   • Research/Investigation — Exploring codebases, understanding systems
   • Architecture/Design — System design, patterns, refactoring
   • Security — Vulnerability assessment, secure coding, audits
   • Operations/DevOps — Deployment, CI/CD, infrastructure, monitoring
   • Data Analysis — Metrics, statistics, analysis, reporting
   • Git/Delivery — Commits, PRs, releases, version management
   • Orchestration/Planning — Task breakdown, delegation, coordination
3. LOAD skills from the Internal Skill Selection Matrix matching the task type.
4. DETECT programming language (if applicable) and load language-specific skills via codebase detection.
5. DELEGATE if complexity warrants (multiple files, architecture decisions, novel problems).

Internal Skill Selection Matrix

External Skill Suggestion (skills.sh)

Suggest an external skill when ALL local options are exhausted and ANY of these conditions are met:

1. Unfamiliar technology — The task involves a library not covered by installed skills.
2. Explicit skill gap — The agent recognises it lacks domain expertise.
3. User signals need — The user asks for help with a specific technology.
4. Repeated uncertainty — 2+ uncertain statements about the same technology in one session.

Guardrails for Suggestions

• Max 1 suggestion per session — Do not nag.
• User consent required — NEVER auto-import.
• 70% confidence threshold — Only suggest when highly confident it helps.
• Max size 5KB — Per system convention.

Execution Rules

1. Classify Context FIRST - Before tools, before thinking, classify the request context.
2. Auto-select Internal Skills - Match keywords from the prompt to the skill matrix.
3. Inject load_skills - Ensure all selected skills are injected into the task call.
4. Identify External Gaps - If local skills are insufficient, check skills.sh (max once).
5. Phase 0 Gate - Prevents proceeding without appropriate skill coverage.

Anti-Patterns

❌ Proceeding without domain-specific skills loaded
❌ Manual skill loading when skill-discovery is possible
❌ Suggesting external skills more than once per session
❌ Auto-importing external skills without explicit user consent
❌ Loading irrelevant skills that waste token context

Integration Points

• Phase 0 gate - Runs before all other processing.
• Skill-auto-loader-config.jsonc - Source of truth for baseline and keyword mappings.
• Universal Skill - Always loaded by default.

KB Reference

~/vaults/baphled/3. Resources/Knowledge Base/AI Development System/Skills/Core-Universal/Skill Discovery.md

Related skills

• agent-discovery — routes to specialist agents; skill-discovery loads domain knowledge
• pre-action — decision framework that benefits from loaded skills