Skill Learning

Skill Learning Methodology

Overview

Transform external knowledge (URLs, files, code) into skill enhancements. Uses novelty-detection to filter training data, matches insights to existing skills, proposes concrete additions.

Core Loop: Source → Extract → Match → Preview → Approve → Apply → Next

Phase 1: Source Processing

URL Sources

OPTIMIZATION: Check for llms.txt first (10x faster when exists)

Detection order: {base_url}/llms-full.txt → llms.txt → llms-small.txt If found: Use directly, skip full page scraping

Primary: WebFetch(url, "Extract technical patterns, gotchas, and implementation details") Fallback: WebFetch("https://r.jina.ai/{url}", ...) if primary blocked

Batch Processing Pattern

When analyzing skill updates across 100+ skills

Strategy: Parallel Read operations for all SKILL.md files in one tool call block Benefit: Claude Code processes parallel independent operations in one response Anti-pattern: Sequential 100+ Read calls (timeout risk)

Example: Read(/skills/skill-1/SKILL.md) Read(/skills/skill-2/SKILL.md) ... Read(/skills/skill-100/SKILL.md) # All in one block

File Sources

Single file: Read(file_path) Simple directory: Glob("*.md", path) + parallel Read Code files: Extract comments, docstrings, error handling patterns

Local Directory Discovery (Plugin/Marketplace Structures)

Step 1: Detect directory structure

Check for common patterns:

• AGENTS.md at root → Parse for skill paths (canonical source)
• plugin.json files → Check for skills/ subdirectory
• Flat skills/*.md → Direct skill files
• Nested skills/*/SKILL.md → Skill subdirectories

Step 2: Discovery commands by structure type

Pattern A: AGENTS.md manifest (preferred)

Read({dir}/AGENTS.md) Parse section → extract relative paths Example paths: "hf-llm-trainer/skills/model-trainer/SKILL.md"

Pattern B: Plugin directories with nested skills

Glob("/skills//SKILL.md", path={dir}) OR Glob("/skills//*.md", path={dir})

Pattern C: Flat skill collection

Glob("skills/*/SKILL.md", path={dir})

Pattern D: Mixed/unknown structure

Glob("**/SKILL.md", path={dir}) # Find all SKILL.md recursively

Step 3: Parallel read all discovered skills

For each discovered path: Read({full_path}) All reads in single block for parallelism

Plugin Directory Example (Illustrative)

Given: ~/.claude/plugins/marketplaces/example-skills/

Step 1: Read AGENTS.md → Discover skill paths:

• plugin-a/skills/database-migrations/SKILL.md
• plugin-b/skills/api-testing/SKILL.md
• plugin-c/skills/docker-compose/SKILL.md

Step 2: Parallel Read all discovered skills

Step 3: For each skill → Extract insights → Match/propose enhancements

Directory Structure Detection Heuristics

Repository Documentation Discovery

When learning from a code repository (not a skills/plugin directory):

Step 1: Identify repo type

Check for indicators:

• package.json → Node.js/TypeScript project
• go.mod → Go project
• Cargo.toml → Rust project
• pyproject.toml → Python project

Step 2: Find documentation files (priority order)

1. README.md, CONTRIBUTING.md, ARCHITECTURE.md (root)
2. docs/.md, documentation/.md
3. src/**/*.md (inline docs)
4. default/content/**/*.json (config/presets)
5. Key source files with heavy comments

Step 3: Find schema/type definitions

• TypeScript: **/*.d.ts, **/types.ts, **/interfaces.ts
• Go: **/*_types.go, **/*_model.go
• JSON Schema: **/*.schema.json
• Validators: **/validator*.js, **/schema*.js

Step 4: Find example/preset files

• **/examples/*, **/presets/*, **/templates/*
• **/default/*, **/samples/*
• **/.example., **/.sample.

Repository Learning Example (Express API)

Given: ~/projects/my-api

Step 1: Detect Node.js project (package.json exists)

Step 2: Read documentation:

• README.md, CONTRIBUTING.md

Step 3: Read schemas/validators:

• src/validators/*.js → Request/response validation schemas
• src/models/*.js → Data model definitions

Step 4: Read examples/presets:

• config/*.json → Environment configurations
• examples/*.json → Sample request/response payloads

Step 5: Read key implementation files:

• src/middleware/auth.js → Authentication flow
• src/middleware/errorHandler.js → Error handling patterns

Step 6: Extract patterns → Create skills:

• express-validation-patterns (from validators)
• api-error-handling (from errorHandler)
• jwt-auth-flow (from auth middleware)

Repo-Specific Extraction Targets

Repo Learning vs Skill Learning

Use Repo Learning when:

• Analyzing a codebase for patterns to adopt
• Extracting schemas/formats (e.g., validation schemas, API specs)
• Learning from reference implementations
• Building NEW skills FROM a repo's patterns

Use Skill/URL Learning when:

• Enhancing EXISTING skills with insights
• Learning from documentation/articles
• Copying skills from plugin marketplaces

Content Cleaning

• Strip navigation, ads, boilerplate
• Preserve code blocks verbatim
• Extract headings as domain signals
• Identify technology keywords (frameworks, libraries, APIs)

Phase 2: Knowledge Extraction

MANDATORY: Apply novelty-detection framework

Skill: novelty-detection

Tier Classification

The Novelty Test

Ask yourself: "Could I have written this WITHOUT reading the source?"

• If YES → Tier 1 (EXCLUDE)
• If NO → Continue to Tier 2-4 classification

Calibration Examples

API Documentation Analysis:

Claim: "OpenAI provides an API for generating text"
→ Tier 1 ❌ — Generic, could write from training data

Claim: "Responses API uses max_output_tokens instead of max_tokens"
→ Tier 2 ✅ — Specific parameter name (HOW)

Claim: "Reasoning models put chain-of-thought in reasoning_content array,
        not content — must sum both for billing"
→ Tier 4 ✅✅✅ — Counter-intuitive, prevents billing surprise

Database Performance:

Claim: "Create indexes on foreign key columns for faster joins"
→ Tier 1 ❌ — Generic DBA advice

Claim: "PostgreSQL partial indexes reduce size 60%, improve write perf 40%"
→ Tier 2 ✅ — Specific feature with quantified benefit

Claim: "Covering indexes avoid heap lookups (3x faster reads, 15% slower writes)"
→ Tier 3 ✅✅ — Quantified trade-off, explains WHY

Claim: "JSONB GIN indexes do NOT support ORDER BY on JSON fields"
→ Tier 4 ✅✅✅ — Contradicts expectation, prevents bug

Framework Patterns:

Claim: "React uses a virtual DOM for efficient updates"
→ Tier 1 ❌ — Training data, everyone knows this

Claim: "Next.js App Router requires 'use client' directive for useState"
→ Tier 2 ✅ — Specific requirement (HOW)

Claim: "Server Components reduce JS bundle by 60% but can't use client state"
→ Tier 3 ✅✅ — Trade-off with quantification (WHY)

Claim: "generateStaticParams runs at BUILD time, not request time —
        dynamic data causes 404s"
→ Tier 4 ✅✅✅ — Contradicts mental model, prevents production bug

Insight Structure

{ "tier": 2, "domain": "sveltekit", "pattern": "Server-only load with +page.server.ts", "insight": "Data fetching in +page.server.ts runs only on server, +page.ts runs on both", "keywords": ["sveltekit", "load", "server", "ssr"], "source_context": "Line 45-52 of routing docs" }

Quality Filter

• Zero Tier 1 leakage (absolute)
• Minimum 3 Tier 2-4 insights per source (or skip)
• Each insight must have domain + keywords

Phase 3: Skill Matching

Discovery

Find all skills

Glob("skills/*/SKILL.md")

Matching Algorithm

1. Exact domain match: Insight domain === skill name (score: 100)
2. Keyword overlap: Insight keywords ∩ skill description/when_to_use (score: 60-90)
3. Technology alignment: Same framework/library family (score: 40-60)
4. No match: Score <40 → propose new skill

Phase 4: Enhancement Proposal

For Each Match (score >= 40)

1. Read current skill Read(skills/{skill-name}/SKILL.md)

2. Identify target section

3. Draft enhancement

• Preserve existing structure exactly
• Add insight in appropriate format for section
• Include source attribution: <!-- Source: {url/file} -->

4. CLEAR Validation Apply skills-enhancer CLEAR framework:

• C: Word count still <5000?
• L: Keywords in right places?
• E: Example shows transformation?
• A: Actionable pattern named?
• R: No duplication, uses references?

Phase 5: User Approval

For Each Enhancement

Present:

1. Skill name being enhanced
2. Insight being added (with tier)
3. Diff preview
4. Word count impact

Ask: "Apply this enhancement? [y/n/edit]"

Response Handling

• y (approve): Apply via Edit tool
• n (reject): Skip, continue to next
• edit: User modifies, then apply

Phase 6: New Skill Proposal

When No Match Found

Insights with no match (score <40):

• Domain: {domain}
• Keywords: {keywords}
• Sample insight: {insight}

Propose new skill? [y/n]

If Approved

Generate using skill-creation methodology: Skill: skill-creation

Phase 7: Loop Control

After Each Source

Summary:

• Insights extracted: X (Tier 2: Y, Tier 3: Z, Tier 4: W)
• Skills enhanced: [list]
• New skills created: [list]
• Rejected: [count]

Next source? (file path, URL, or 'done')

Quality Gates

Absolute Rules

• Zero Tier 1 insights in skills
• User approves each change (no auto-apply)
• Diff preview shown before any edit
• Source attribution in comments

Warning Triggers

• Skill exceeds 5000 words → suggest splitting
• Large source (10K+ pages) → create router skill
• Insight duplicates existing content → skip
• CLEAR validation fails → revise before applying

Quick Reference