Meta agent skills

Meta-Agent Skills

Purpose

This skill serves as a "Meta-Skill" that bootstraps the Agentic Makefile environment. It empowers the Agent to analyze the repository's technology stack (e.g., Python/Poetry, Node/Next.js, Go), detect the AI environment (Claude Code, Cursor, Gemini), and generate specialized, ready-to-use Agent Skills and Subagents.

When to Use

• When initializing a new repository for AI agent use.
• When the technology stack changes (e.g., switching from Pip to Poetry).
• When you want to reset or update the standard agent capabilities.
• When asked to "setup skills", "maintain agent rules", or "install standard agents".

Instructions

1. Detect AI Environment:

   • Check for .claude/, .cursor/, or .gemini/ directories to determine the target AI platform.
   • Claude Code Detection: If .claude/ exists, Claude Code is a primary target.
     • Prefix Selection: Claude Code does not support recursive search for skills/agents. You MUST use a flat structure with a prefix to identify generated components (e.g., ma-, meta-, m-).
     • User Consultation: Present 3-5 candidate prefixes (e.g., ma-, meta-agent-, m-, agent-, gen-) and ask the user to choose one or provide their own.
   • Cursor Detection: If .cursor/ exists, Cursor is a primary target. Cursor supports recursive search.
   • Default: If ambiguous, prioritize .claude/ as the standard, following the flat structure protocol.

2. Analyze Codebase:

   • Review Documentation: Read README.md, CONTRIBUTING.md, DEVELOPMENT.md, or other relevant documentation to understand the project structure, development workflows, and any specific commands recommended for the codebase.
   • Detect Sub-Projects: Recursively search for "logical project boundaries" in sub-directories. Look for files like package.json (Node.js), go.mod (Go), pyproject.toml or requirements.txt (Python), main.tf or *.tf (Terraform), etc.
   • Detect Multi-Layered Builds: Search for files that indicate a layered build or deployment process, such as Dockerfile, docker-compose.yml, Earthfile, Tiltfile, Skaffold.yaml, or kustomization.yaml.
   • Map Tech Stack per Project: For each detected sub-project, determine its specific tech stack and how to run builds, linters, and tests within its directory.
   • Analyze Layered Commands: Categorize commands into logical layers (e.g., App for compilation, Docker for image building, Infra for deployment or local orchestration).
   • Identify Test Types: Look for tests/unit, tests/integration, cypress, playwright, etc., to distinguish between Unit, Integration, and E2E tests for each project.
   • Identify Security Tools: Check if trivy, osv-scanner, or other security tools are configured or available.
   • Identify Setup Scripts: Look for pre-commit config, Makefile, or setup scripts to include in setup-dev-env.

3. Verify Commands:

   • Before generating skills, proactively verify that the detected commands work in their respective project environments.
   • Run command --help, command --version, or similar check for each primary command in the correct working directory.
   • If a command fails or is missing, investigate alternatives or suggest installation in the final report.

4. Generate Skills & Agents:

   • Read the templates located in assets/templates/skills/ and assets/templates/agents/.
   • Strict Policy: You MUST NOT generate any subagent or Agent Skill if its corresponding template does not exist in assets/templates/agents/ or assets/templates/skills/.
   • Instantiate Templates:
     • For each skill template, populate the Commands table with the verified commands for all detected sub-projects.
     • Build Separation: Distinguish between project compilation (App layer) and container image building (Docker layer).
       • Use build-project template for compilation/build commands (e.g., npm run build, go build).
       • Use build-container-image template for containerization commands (e.g., docker build, earthly --push +docker).
     • Each row in the table MUST include the Order, Component, Path (relative to root), Layer (e.g., App, Docker), Command, and Description.
     • Ensure the order of commands is logical (e.g., compile app before building docker image).
   • Write the generated files to the target directory based on the platform:
     • Claude Code (Flat Structure):
       • Skills: .claude/skills/<prefix><skill-name>/SKILL.md (e.g., .claude/skills/ma-lint-fix/SKILL.md).
       • Agents: .claude/agents/<prefix><agent-name>.md (e.g., .claude/agents/ma-maintainer-agent.md).
     • Cursor (Nested Structure):
       • Skills: .cursor/skills/meta-agent-skills/<skill-name>/SKILL.md (e.g., .cursor/skills/meta-agent-skills/lint-fix/SKILL.md).
       • Agents: .cursor/agents/meta-agent-skills/<agent-name>.md (e.g., .cursor/agents/meta-agent-skills/maintainer-agent.md).
   • Bind Skills to Agents:
     • For each generated agent, identify the skills required from its template frontmatter.
     • Synchronize the Capabilities section between <!-- SKILLS_START --> and <!-- SKILLS_END --> markers.
     • Link Resolution:
       • Claude Code: Use links like [lint-fix](../skills/<prefix>lint-fix/SKILL.md).
       • Cursor: Use links like [lint-fix](../../skills/meta-agent-skills/lint-fix/SKILL.md).
     • Ensure each mentioned skill is linked to its respective SKILL.md file.
   • Note: For test-* skills, only generate the ones that match the detected test types.

5. Verify & Fix Generated Output:

   • Audit: Read a sample of the generated SKILL.md files (prioritize lint-fix and build-project).
   • Verify Templates: Verify that every generated subagent and Agent Skill has a corresponding template in the assets directory. If you find any generated file that does not have a corresponding template, you MUST delete it.
   • Check for Placeholders: Ensure no unpopulated templates like {{ command }} remain in the generated files.
   • Path Validation: Verify that the Working Directory paths specified in the tables actually exist relative to the workspace root.
   • Immediate Remediation: If errors, broken links, or missing information are found, use editing tools to fix the generated files immediately.

6. Execute Generated Skills & Agents:

   • Smoke Test: Execute a subset of the generated skills to verify their real-world functionality.
   • Priority Skills: Run setup-dev-env (if applicable), followed by lint-fix, build-project, and build-container-image.
   • Verify Subagents: If a subagent was generated, consider invoking it for a simple query (e.g., "Analyze the current state of the codebase").
   • Error Handling: If execution fails, analyze the output, fix the generated skill/agent, and re-run until successful.

7. Report:

   • List the skills and agents created.
   • Mention which stack and test types were detected.
   • Report the results of command verification (which commands are confirmed and which might need setup).
   • Report on the Verification & Fix results (e.g., "Verified all generated skills; fixed 1 path error in lint-fix").
   • Report on the Execution results (e.g., "Successfully ran lint-fix, build-project, and build-container-image skills").

Capabilities Generated

• Core Skills: lint-fix (includes type checking), build-project, build-container-image, update-deps, docs-gen-readme, security-scan, setup-dev-env, add-skill-templates, add-agent-templates, mend-agent-templates.
• Test Skills: test-unit, test-integration, test-e2e.
• Subagents: codebase-maintainer-agent, security-auditor-agent, qa-engineer-agent, template-factory-agent.

References

• Agent Skills Standard
• Claude Code
• Cursor
• Gemini CLI
• Codex