This repository installs AI agent skills that influence how coding agents reason, read files, run commands, and make changes. Please read this before installing on any sensitive machine.

Upstream skills in curated-skills.json are pinned to full 40-character SHA commit hashes — not branch names or tags. This means you always get the exact code that was reviewed, not whatever happens to be at HEAD of an upstream repo.

{
  "repo": "mattpocock/skills",
  "commit": "9f2e0bd0ea776eb6372eb81fa8a4a47814a8404a",
  "verified_date": "2026-05-11"
}

To audit whether upstream repos have moved since the last pin:

node scripts/audit-upstream.mjs

This script reports UP_TO_DATE, BEHIND, or UNKNOWN for each upstream. It never auto-updates. Pin bumps require a human decision.

The gate runner enforces a protected_paths list from your .buildloop.yml. Any AI agent action that touches a protected path (secrets, CI workflows, infra configs, migration files) is blocked and reported before it executes.

protected_paths:
  - ".env*"
  - "**/*.key"
  - "**/*.pem"
  - ".github/workflows/**"

You control this list. Add your project's sensitive paths before enabling autonomous execution.

• Read scripts/install.ps1 (Windows) or scripts/install.sh (macOS/Linux) before running any remote one-line command on a sensitive machine.
• Prefer a local clone install when working on regulated, financial, health, government, or production systems.
• Review each installed SKILL.md before enabling broad agent autonomy.

• Copies the local wrapper skills from this repository.
• Downloads curated upstream skills from their original public GitHub repositories at pinned commits.
• When run as a one-line remote installer, downloads the Buildloop payload to a temporary directory so local and upstream skills can both be installed.
• For Claude Code only, installs convenience command aliases in ~/.claude/commands so users can run /orchestrator or /buildloop.
• Removes one non-standard frontmatter key from imported skills for agent compatibility.
• Does not request secrets, credentials, API keys, or production access.
• Does not modify project source files. It writes only to the selected agent's skills directory and, for Claude Code, the user's Claude command directory.

Do not open a public GitHub issue for security vulnerabilities. Public issues disclose the problem to attackers before a fix is available.

Use GitHub's built-in private reporting:

1. Go to the Security tab of this repository.
2. Click "Report a vulnerability".
3. Fill in the details privately. Only maintainers can see this report.

If GitHub private reporting is unavailable or you prefer email, contact:

security@mithunyc.dev

Include:

• A description of the vulnerability and its potential impact.
• Steps to reproduce (if applicable).
• Any suggested fix or mitigation.

You will receive a response within 5 business days. If you do not hear back, re-send with [SECURITY] in the subject line.

For installer failures, agent compatibility issues, or unexpected behavior — open a standard GitHub issue with:

• Target agent: Codex, Claude Code, Cursor, or Antigravity.
• OS and shell.
• Install command used.
• Full error output.
• Whether the skill appears after restarting the agent.

This is a public skill distribution. No skill in this repo should be granted production database access, secrets, or infrastructure permissions by default. Skills are reasoning guides, not privileged processes.

If you choose to run an agent with access to production systems, apply the principle of least privilege: give the agent only the permissions it needs for the specific task, revoke them when done, and audit the gate-results.json log before merging any changes.