INSTRUCTIONS FOR CLAUDE: Skill Learning Methodology

Quick Reference

Intent	Action
Extract patterns from a URL	Use `WebFetch` and Phase 1a
Extract patterns from a codebase	Use `Glob`/`Read` and Phase 1b
Match an extracted insight to a skill	Execute Phase 3 Scoring Algorithm
Propose an enhancement to the user	Follow the format in Phase 5
Apply an approved enhancement	Immediately `Edit` the target `SKILL.md`

When to Use

User provides a URL to documentation or an article and asks to learn from it.
User points to a local directory or repository to extract design patterns.
User provides a path to another agent's skill directory or a marketplace.
Proactively when a user pastes a large block of reference code and asks to "save this pattern."

System Prompt

You are executing a rigid 7-phase knowledge extraction protocol. Your primary objective is to extract only novel technical patterns from external sources and inject them into the user's local skills/ directory.

CRITICAL DIRECTIVE: You have a natural tendency to summarize everything you read. DO NOT DO THIS. The user's context window is precious. You must aggressively filter out "Tier 1" knowledge (things you already know from your pre-training data) and only retain Tier 2, 3, or 4 insights.

Core Execution Loop: 1. Source → 2. Extract → 3. Match → 4. Preview → 5. Approve → 6. Apply → 7. Loop

Anti-Patterns

Anti-Pattern	Problem	Fix
Summarizing Training Data	Bloats the context window with useless Tier 1 facts (e.g. "React uses a Virtual DOM").	Ruthlessly apply the Novelty Test. Exclude Tier 1.
Sequential File Reading	Calling `Read()` 100 times in a loop will cause the model to time out.	Use Parallel Tool Calls inside a single block.
Asking before Editing	If the user already said "Apply" in Phase 5, pausing to ask permission to use the `Edit` tool is maddening.	Execute the `Edit` tool immediately upon user approval.
Missing Source Links	Future agents won't know where the pattern came from.	Always append `<!-- Source: {url/file} -->`.

Phase 1: Source Processing (Execution Steps)

1a. URL Sources

ACTION: You must attempt to fetch the URL using the WebFetch tool.

OPTIMIZATION: Check for llms.txt first. Attempt {base_url}/llms-full.txt → llms.txt → llms-small.txt. If found, use it directly to avoid scraping HTML.
Primary: WebFetch(url, format="markdown")
Fallback: If blocked by bot-protection, use WebFetch("https://r.jina.ai/{url}").

1b. File Sources & Batch Processing

ACTION: If the source is a local directory or repository, use Glob and Read.

Strategy: When analyzing multiple files (e.g., discovering existing skills), you MUST use Parallel Tool Calls. Output all your Read tool calls in a single response block.
Anti-pattern: Never use sequential Read calls for a large list of files. You will time out.

1c. Local Directory Discovery (Plugin/Marketplace Structures)

ACTION: Use Glob and Read tools to find existing SKILL.md files or AGENTS.md manifests.

Structure A: AGENTS.md manifest (preferred) Read {dir}/AGENTS.md. Parse <available_skills> section and extract relative paths.
Structure B: Plugin directories with nested skills Use Glob with pattern */skills/*/SKILL.md or */skills/*/*.md.
Structure C: Flat skill collection Use Glob with pattern skills/*/SKILL.md.
Structure D: Mixed/unknown structure Use Glob with pattern **/SKILL.md to find all skills recursively.

Once discovered, execute Parallel Read on all discovered skills.

1d. Repository Documentation Discovery

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

Identify repo type: Look for package.json, go.mod, Cargo.toml, pyproject.toml.
Find docs: README.md, CONTRIBUTING.md, docs/*.md.
Find schemas/types: **/*.d.ts, **/*_types.go, **/*.schema.json.
Find examples: **/examples/*, **/presets/*.

Use Repo Learning when: Extracting schemas/formats, learning from reference implementations, or building NEW skills FROM a repo's patterns. Use Skill/URL Learning when: Enhancing EXISTING skills with insights from documentation/articles or copying skills from marketplaces.

1e. Content Cleaning

Before processing, mentally strip navigation, ads, and boilerplate. Preserve code blocks verbatim.

Phase 2: Knowledge Extraction

MANDATORY: You must apply the novelty-detection framework to filter the extracted content.

Tier Classification

You must classify every extracted insight into one of four tiers:

Tier	Include?	Signal
1	EXCLUDE	Could write without source (training data)
2	Include	Shows HOW (implementation-specific)
3	High value	Explains WHY (architectural trade-offs)
4	Highest	Contradicts assumptions (counter-intuitive)

The Novelty Test

For every insight, ask yourself: "Could I have written this WITHOUT reading the source?"

If YES → It is Tier 1. You MUST EXCLUDE IT.
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 Requirements

You must structure each extracted insight logically before scoring it:

{
  "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 the source).
Each insight must have a domain + keywords.

Phase 3: Skill Matching

3a. Discovery

ACTION: Find all existing skills using the Glob tool with the pattern skills/*/SKILL.md. If you already discovered them in Phase 1, you can skip this step.

3b. Matching Algorithm

You must score each extracted insight against the user's existing skills to find the best home for it.

Exact domain match: Insight domain === skill name (score: 100)
Keyword overlap: Insight keywords ∩ skill description (score: 60-90)
Technology alignment: Same framework/library family (score: 40-60)
No match: Score <40 → Skip enhancement and propose a new skill instead.

Phase 4: Enhancement Proposal

For Each Match (score >= 40)

1. Read current skill ACTION: Use the Read tool to read the contents of skills/{skill-name}/SKILL.md.

2. Identify target section Determine where the insight fits inside the existing skill file:

Insight Type	Target Section
Quick fact	Quick Reference table
Pattern + example	Patterns / Examples
Gotcha / warning	Anti-Patterns / Common Mistakes
Workflow step	Process / Workflow
Validation rule	Checklist

3. Draft the enhancement

Preserve the existing structure exactly.
Add the insight in the appropriate format for that section.
You MUST include source attribution:

4. CLEAR Validation Mentally apply the CLEAR framework to your proposed addition:

Compact: Will the word count stay under 5000?
Labeled: Are keywords in the right places?
Example: Does the example show a clear transformation?
Actionable: Is the actionable pattern named?
References: Is there no duplication, and does it use references?

Phase 5: User Approval

The Proposal Format

For each valid enhancement, you must present the proposal to the user exactly like the examples shown in the Examples section at the bottom of this document.

Present the:

Target Skill name
Insight summary and Tier
Diff preview of what you are going to add
Source attribution

ACTION: Use the AskUserQuestion tool to ask: "Apply this enhancement?" with the options: Apply, Skip, or Edit.

Response Handling

Apply: Proceed to Phase 6.
Skip: Skip to the next candidate.
Edit: User modifies the text, then you proceed to Phase 6.

Phase 6: Apply & New Skill Proposal

6a. Apply Enhancement

ACTION: If the user selected 'Apply', you MUST immediately use the Edit tool on skills/{skill-name}/SKILL.md to insert the drafted block into the target section. Do not just say you will do it, execute the tool.

6b. When No Match Found (New Skills)

For insights with no match (score <40), present the user with a summary of the domain and keywords. Ask: "Propose new skill for {domain}?" [y/n] If approved, generate the new skill directory and SKILL.md file using the Write tool.

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

Examples

Example 1: Extracting from Documentation

Source: User runs /learn https://svelte.dev/docs/kit/state-management Insight (Tier 3): SvelteKit 5 replaces store subscriptions with runes ($state, $derived) for reactivity. Target Skill: sveltekit-patterns Agent Output Preview:

## Enhancement Proposal (Score: 85, Tier: 3)

**Insight**: SvelteKit 5 relies on runes ($state, $derived) instead of store subscriptions for reactive UI state.
**Target Skill**: skills/sveltekit-patterns/SKILL.md
**Section**: Patterns

**Proposed Addition**:
### Runes vs Stores (Svelte 5)
<!-- Source: https://svelte.dev/docs/kit/state-management -->
Replace old `writable` stores with `$state()` runes for component-level reactivity. Do not use `$:` for derived state; use `$derived()` instead.

Apply this enhancement? [y/n/edit]

Example 2: Extracting from Codebase

Source: User runs /learn ~/projects/api-gateway Insight (Tier 2): The codebase uses a specific Zod schema wrapper (validateRequest(schema)) for all incoming request validation. Target Skill: api-middleware Agent Output Preview:

## Enhancement Proposal (Score: 70, Tier: 2)

**Insight**: Local convention uses `validateRequest(schema)` middleware pattern for all Zod validation.
**Target Skill**: skills/api-middleware/SKILL.md
**Section**: Anti-Patterns / Common Mistakes

**Proposed Addition**:
### Zod Validation Convention
<!-- Source: ~/projects/api-gateway/src/middleware/ -->
**Do NOT** validate Zod schemas inline inside the route handler. 
**DO** use the `validateRequest(schema)` middleware exported from `src/middleware/validator.ts`.

Apply this enhancement? [y/n/edit]

skill-learning