/learn

name: learn
description: Provides autonomous project pattern learning by analyzing the codebase to discover development conventions, architectural patterns, and coding standards, then generates project rule files in .claude/rules/. Use when user asks to "learn from project", "extract project rules", "analyze codebase conventions", "discover project patterns", or wants to auto-generate Claude Code rules for the current project.
allowed-tools: Read, Write, Edit, Bash, Glob, Grep, Task, AskUserQuestion

Learn

Autonomously analyzes a project's codebase to discover development patterns, conventions, and architectural decisions, then generates project rule files in `.claude/rules/` for Claude Code to follow.

Overview

This skill acts as the **Orchestrator** in a two-agent architecture. It coordinates the overall workflow: gathering project context, delegating deep analysis to the `learn-analyst` sub-agent, filtering and ranking results, presenting findings to the user, and persisting approved rules to `.claude/rules/`.

The separation of concerns ensures the analyst operates with a focused forensic prompt while the orchestrator manages user interaction and file persistence.

When to Use

Use this skill when:

• User asks to "learn from this project" or "understand project conventions"
• User wants to auto-generate `.claude/rules/` files from the existing codebase
• User asks to "extract project rules" or "discover patterns"
• User wants Claude Code to learn the project's coding standards
• After joining a new project and wanting to codify existing conventions
• Before starting a large feature to ensure Claude follows project patterns

**Trigger phrases:** "learn from project", "extract rules", "analyze conventions", "discover patterns", "generate project rules", "learn codebase", "auto-generate rules"

Instructions

Phase 1: Project Context Assessment

Before delegating to the analyst, gather high-level project context:

1. **Verify project root**: Confirm the current working directory is a project root (has `package.json`, `pom.xml`, `pyproject.toml`, `go.mod`, `.git/`, or similar markers)

2. **Check existing rules**: Scan for pre-existing rule files to understand what is already documented:

# Check for existing rules
ls -la .claude/rules/ 2>/dev/null || echo "No .claude/rules/ directory found"
cat CLAUDE.md 2>/dev/null || echo "No CLAUDE.md found"
cat AGENTS.md 2>/dev/null || echo "No AGENTS.md found"
ls -la .cursorrules 2>/dev/null || echo "No .cursorrules found"

3. **Assess project size**: Get a quick overview of the project scope:

# Quick project overview
find . -maxdepth 1 -type f -name "*.json" -o -name "*.toml" -o -name "*.xml" -o -name "*.gradle*" -o -name "Makefile" -o -name "*.yaml" -o -name "*.yml" | head -20
find . -type f -name "*.ts" -o -name "*.js" -o -name "*.java" -o -name "*.py" -o -name "*.go" -o -name "*.php" | wc -l

4. **Inform the user**: Briefly tell the user what you found and that you are about to start analysis:

• "I found a [TypeScript/NestJS] project with [N] source files and [M] existing rules. Starting deep analysis..."

Phase 2: Delegate to Analyst Sub-Agent

Invoke the `learn-analyst` sub-agent to perform the deep codebase analysis.

Use the **Task tool** to delegate analysis to the `learn-analyst` agent:

• **Agent**: `learn-analyst`
• **Prompt**: "Analyze the codebase in the current working directory. Follow your full process: discovery, pattern extraction, classification, and prioritization. Return your findings as a JSON report."
• **Mode**: Run synchronously to receive the JSON report directly

The analyst will return a structured JSON report with classified findings.

Phase 3: Review and Filter Results

Process the analyst's report:

1. **Parse the JSON report** returned by the analyst 2. **Validate findings**: Ensure each finding has:

• A clear title
• Evidence from at least 2 files
• Impact score ≥ 4 (discard low-impact findings)
• Well-formed markdown content

3. **Deduplicate against existing rules**: Compare each finding title and content against existing `.claude/rules/` files. Skip findings that duplicate existing rules. 4. **Select top 3**: From the remaining findings, select the top 3 by impact score. If fewer than 3 remain after filtering, present whatever is left. 5. **If zero findings remain**: Inform the user that the project is already well-documented or no significant undocumented patterns were found.

Phase 4: Present to User

Present the filtered findings to the user in a clear, structured format:

I analyzed your codebase and found N patterns worth documenting as project rules:

1. **[RULE]** <Title> (Impact: X/10)
   <One-line explanation>

2. **[RULE]** <Title> (Impact: X/10)
   <One-line explanation>

3. **[RULE]** <Title> (Impact: X/10)
   <One-line explanation>

Then ask the user for confirmation using **AskUserQuestion**:

• Present choices: "Save all N rules", "Let me choose which ones to save", "Cancel — don't save anything"
• If the user wants to select individually, present each rule one by one with "Save / Skip" options
• **Never save automatically** — always require explicit user approval

Phase 5: Persist Approved Rules

For each approved rule:

1. **Ensure directory exists**:

mkdir -p .claude/rules

2. **Generate the file name**: Use the finding's `title` field converted to kebab-case:

• Example: `"API Response Envelope Convention"` → `api-response-envelope-convention.md`
• Avoid generic names like `rule-1.md` or `learned-pattern.md`

3. **Check for conflicts**: Before writing, check if a file with the same name already exists:

• If it exists, present a diff to the user and ask whether to replace, merge, or skip

4. **Write the rule file**: Create the file in `.claude/rules/` with the analyst's pre-formatted content

5. **Confirm to user**: After saving, list all created files:

✅ Rules saved successfully: