claude-code-superpowers

Five components extracted from a real-world Claude Code setup that has been used daily for months. None of them are bound to a specific business domain - they are the parts that turned out to be reusable.

What's in the box

Component	What it does	Where
session-bootstrap hook	Reads STATUS.md, the most recent active PROJECT.md, the latest briefing, and the tail of your wiki log at every session start. The model walks in already knowing what you're working on.	hooks/session-bootstrap.py
claude-doctor skill	Read-only health check: MCP servers reachable, every SKILL.md valid, memory index parseable, CLAUDE.md references exist, scheduled tasks alive, wiki not stale. Returns green / yellow / red with concrete fixes.	skills/claude-doctor/
skills-supply-chain-audit skill	Read-only security scanner over your skill directories. Catches new external URLs, curl | bash patterns, npx/uvx runs, eval(), base64 payloads, credential leaks, plus per-project hash drift against a lockfile. Runs on demand or in a cron.	skills/skills-supply-chain-audit/
quality-gate skill	Two-mode quality gate for any text the model produces in your name. Mode 1 - deterministic regex lint (em-dashes, AI cliches, anti-thesis patterns). Mode 2 - three parallel LLM sub-agents evaluating clarity, facts, and tonal consistency.	skills/quality-gate/
auto-memory pattern	Schema, four type definitions, and four generic example memory files. The model already supports file-based memory; this directory shows how to structure it so it survives across sessions.	memory/

Plus one bonus pattern doc:

multi-agent-batch-orchestration	When and how to fire 3-5 sub-agents in parallel from inside a single skill. Includes a demo prompt to test the pattern.	patterns/

How to install (read this part)

Don't git clone and copy files into your setup. A lot of this code makes assumptions about your project layout, your operating system, and what you already have installed. Half of it might not fit.

Instead, in your own Claude Code session, paste this prompt:

Read the repo https://github.com/Luis247911/claude-code-superpowers - the README first, then the individual component folders under skills/, hooks/, memory/, patterns/.

Before you suggest installing anything:

1. Scan my current Claude Code setup: ~/.claude/skills/, ~/.claude/hooks/, ~/.claude/settings.json, and any CLAUDE.md in the current project. Tell me in one paragraph what you see.

2. For each component in the repo, compare:
   a) Do I already have something equivalent?
   b) If yes - is mine better, worse, or just oriented differently?
   c) If no - what concrete problem would the repo component solve for me?

3. Give me a prioritised recommendation list. Per component:
   - "adopt" / "skip" / "selectively port" with one line of reasoning.
   - If "selectively port": which file(s) and what adaptations.

4. Do not auto-download or clone the repo. Read the files via web fetch or ask me to show you specific paths. Only add code to my setup after I explicitly approve a step from the list above.

This treats the repo as a reference, not a payload. The model assesses fit before touching your system.

Why this format

Skills, hooks, and memory in Claude Code are powerful precisely because they're personal. A copy-pasted skill from someone else's setup either doesn't fire (because it assumes a directory you don't have) or fires too often (because the trigger words match unrelated work). Letting the model do the diff between this repo and your setup gives you a much higher hit rate.

Configuration knobs (per component)

Each component is parametrised via environment variables. Read the SKILL.md inside each folder for the full list. The most important ones:

• CLAUDE_BOOTSTRAP_ROOT - point the bootstrap hook at your project (default: current working directory).
• CLAUDE_DOCTOR_PROJECT_ROOT - point doctor at your project.
• CLAUDE_DOCTOR_TASK_PREFIXES - comma-separated substrings if you want doctor to verify scheduled tasks (Windows only; empty = skip).
• CLAUDE_AUDIT_PROJECT_ROOT - point the audit at your project.
• ~/.claude/supply-chain-allow-list.json - your trusted hosts for the audit (an example file is in skills/skills-supply-chain-audit/).

Known limitations

• Language: quality-gate/lint.py is tuned for German output (cliche list, anti-thesis patterns). The structure is generic; replace the word lists for English.
• OS: claude-doctor checks scheduled tasks via Windows schtasks. On macOS / Linux that section is skipped; cron / launchd checks are not implemented yet.
• Hook integration: The session-bootstrap hook needs to be registered in your ~/.claude/settings.json. See hooks/settings.example.json for the snippet.
• Memory directory: Claude Code creates ~/.claude/projects/<hash>/memory/ automatically the first time it stores something. The hash is derived from your absolute project path.

License

MIT. See LICENSE.

Contributing

These components grew out of one person's setup and were extracted once they had stabilised. If you adapt them and find a better default, open a PR - especially for the platform-specific parts (cron / launchd in doctor, English word lists in quality-gate).