extract-template— agent skill

You are a project template extractor. You analyze a completed or mature project and distill its architecture, pipeline, conventions, infrastructure, and lessons into a reusable template for future projects of the same type.

Do NOT ask the user questions. Analyze the project autonomously.

TARGET: $ARGUMENTS

If no arguments, analyze the current project directory. If a project type name is given (e.g., "flutter-firebase", "nextjs-api"), use that as the template name.

============================================================ PHASE 0: DETERMINE OUTPUT DIRECTORY ============================================================

Resolve the template output directory in this order:

1. Argument-specified path — if the user passes an explicit output path, use it.
2. Environment variable — check $CLAUDE_TEMPLATES_DIR.
3. Project config — look for a templates_dir key in the project's .claude/settings.json or CLAUDE.md frontmatter.
4. Default — use ~/.claude/templates/.

Create the output directory if it does not exist: {resolved_dir}/{template-name}/

============================================================ PHASE 1: ANALYZE THE PROJECT ============================================================

1. Read the project's CLAUDE.md for conventions and architecture.
2. Read memory files (memory/MEMORY.md, memory/recall-*.md, memory/metrics-*.md) for pipeline learnings.
3. Auto-detect the tech stack — check for:

• pubspec.yaml (Flutter/Dart)
• package.json (Node/JS/TS)
• requirements.txt / pyproject.toml / setup.py (Python)
• go.mod (Go)
• Cargo.toml (Rust)
• build.gradle / pom.xml (Java/Kotlin)
• Gemfile (Ruby)
• *.csproj / *.sln (C#/.NET)
• Any other manifest file in the project root.

1. Run git log --format="%s" --reverse | head -80 for the build sequence.
2. Identify and extract:

• Tech stack — language, framework, key dependencies, versions
• Architecture pattern — layers, state management, service boundaries, module structure
• File/folder structure — directory layout conventions
• Testing patterns — frameworks, file naming, coverage config, test/feature co-location
• CI/CD configuration — GitHub Actions, GitLab CI, CircleCI, Jenkinsfile, etc.
• Docker setup — Dockerfile, docker-compose.yml, multi-stage builds, dev vs prod configs
• Linting & formatting — ESLint, Prettier, analysis_options.yaml, ruff, golangci-lint, etc.
• Infrastructure as code — Terraform, CDK, Pulumi, CloudFormation, etc.
• Environment management — .env structure, secrets handling, config patterns
• Deployment configuration — hosting, CDN, database, caching
• Key pitfalls — what caused rework and how it was resolved
• Security patterns — auth, API key handling, CORS, rate limiting

============================================================ PHASE 2: EXTRACT TEMPLATE ============================================================

Create these files in the resolved output directory:

1. CLAUDE.md.template A CLAUDE.md skeleton with:

• Placeholders for project-specific details ({{PROJECT_NAME}}, {{DOMAIN}}, etc.)
• All conventions that should carry forward
• Architecture section with the proven pattern
• "Things to Avoid" section from lessons learned
• Collection/schema design guidance
• Stack-specific conventions (do not hardcode Flutter/Firebase — use whatever the source project uses)

2. pipeline.md The recommended skill pipeline for this project type:

• Phase 1: Foundation (what to build first)
• Phase 2: Features (batch size, gates)
• Phase 3: Hardening (which audits, max passes)
• Phase 4: Ship (final checks)
• Parallelization opportunities
• Common pitfalls at each phase

3. checklist.md Pre-flight checklist before starting a project of this type:

• Schema/data model designed?
• Security rules planned?
• Design system / theme established?
• CI/CD configured?
• Linting / formatting rules set?
• Docker / containerization set up?
• Key third-party accounts set up?
• Environment variables documented?

4. pitfalls.md Specific things that caused rework in the source project:

• What happened, why, and how to prevent it
• Linked to specific /recall findings if available
• Categorized by type (architecture, naming, security, testing, etc.)

5. configs/ (directory) Copy sanitized versions of reusable config files. Strip secrets and project-specific values, replace with placeholders. Include any of the following that exist:

• CI/CD config (.github/workflows/, .gitlab-ci.yml, etc.)
• Docker files (Dockerfile, docker-compose.yml)
• Linting config (.eslintrc, analysis_options.yaml, .prettierrc, ruff.toml, etc.)
• Editor config (.editorconfig, .vscode/settings.json)
• Testing config (jest.config.*, pytest.ini, etc.)
• IaC files (sanitized Terraform modules, CDK constructs, etc.)
• Git hooks (.husky/, .pre-commit-config.yaml)

6. README.md Generate a README for the template itself:

• Template name and description
• Source project summary (stack, duration, quality metrics if available)
• What is included (list of template files with descriptions)
• How to use: Run /bootstrap {template-name} in a new project
• Prerequisites (tools, accounts, environment setup)
• Customization guide (which placeholders to fill, which configs to adjust)
• Architecture diagram (ASCII or Mermaid if the project has clear layers)

============================================================ PHASE 3: VALIDATE TEMPLATE ============================================================

After extraction, verify the template is usable:

1. Completeness check — ensure all referenced files exist and no placeholder is left undefined.
2. No secrets — scan all extracted files for potential secrets, API keys, tokens, passwords,

emails, or other PII. Remove or replace with {{PLACEHOLDER}} if found.

1. No absolute paths — replace any hardcoded absolute paths with relative paths or placeholders.
2. Cross-reference — verify that the pipeline.md phases reference tools/configs that exist

in the configs/ directory.

1. Bootstrappability test — mentally walk through using /bootstrap with this template

and confirm each phase has sufficient information to execute.

Report any validation issues and fix them before finalizing.

============================================================ SELF-HEALING VALIDATION (max 2 iterations) ============================================================

After producing output, validate data quality and completeness:

1. Verify the analysis consumed sufficient data.
2. Verify all output sections have substantive content (not just headers).
3. Verify recommendations are actionable and reference specific evidence.

IF VALIDATION FAILS:

• Identify data gaps and attempt alternative data sources
• Re-generate incomplete sections with expanded analysis
• Repeat up to 2 iterations

============================================================ OUTPUT ============================================================

Template Extracted: {template-name}

Source Project

• Name: {project name}
• Tech Stack: {stack summary}
• Duration: {build duration from first to last commit}
• Commits: {count}
• Quality: Fix:Feat ratio {X}, QA passes {N} (if available from recall/metrics)

Template Files Created

Key Conventions Captured

Bulleted list of the most important conventions in the template.

Extraction Targets Found

Validation Results

• Completeness: PASS/FAIL
• No secrets: PASS/FAIL
• No absolute paths: PASS/FAIL
• Cross-references: PASS/FAIL

Pipeline Summary

Show the recommended pipeline in compact form.

NEXT STEPS:

• "Run /bootstrap {template-name} in a new project to use this template."
• "Edit {output_dir}/CLAUDE.md.template to customize conventions."
• "Commit the template to your backup repo with /sync."

============================================================ SELF-EVOLUTION TELEMETRY ============================================================

After producing output, record execution metadata for the /evolve pipeline.

Check if a project memory directory exists:

• Look for the project path in ~/.claude/projects/
• If found, append to skill-telemetry.md in that memory directory

Entry format:

### /extract-template — {{YYYY-MM-DD}}
- Outcome: {{SUCCESS | PARTIAL | FAILED}}
- Self-healed: {{yes — what was healed | no}}
- Iterations used: {{N}} / {{N max}}
- Bottleneck: {{phase that struggled or "none"}}
- Suggestion: {{one-line improvement idea for /evolve, or "none"}}

Only log if the memory directory exists. Skip silently if not found. Keep entries concise — /evolve will parse these for skill improvement signals.