superpowers-plus

AI coding assistants skip the practices that keep bugs out of production. They implement the first idea without weighing alternatives, patch symptoms instead of finding root causes, and declare work done without verifying it. Asking them to behave better in a system prompt doesn't hold: under context pressure, the instructions get forgotten.

superpowers-plus is 124 skills that make an assistant follow those practices, plus the machinery that enforces them from outside the model. Skills give the assistant a procedure: reproduce before fixing, generate three designs before picking one, send a diff to parallel specialist reviewers. Lifecycle hooks and git commit gates check the result. Git runs those gates whether or not the assistant remembers to ask for review.

It builds on Jesse Vincent's obra/superpowers. Superpowers teaches an agent how to work; superpowers-plus adds more skills and the hooks and gates that make the work hard to skip. It also covers non-coding work: documents, wikis, issue tracking, research.

Quick Start

Requires bash 4+, git, and Node.js 18+. macOS ships bash 3.2, so run brew install bash first.

git clone https://github.com/bordenet/superpowers-plus.git && cd superpowers-plus
bash install.sh                 # skills, hooks, and runtime
bash tools/install-hooks.sh     # git commit and push gates

Then tell your assistant what you're doing:

Full install options (Claude Code plugin, Codex, OpenCode, MCP server, Windows/WSL) are in docs/INSTALLATION.md.

What Changes

"This test started failing after yesterday's change. Fix it."

Without:  edit code -> test passes -> "Done."
With:     reproduce -> find the root cause -> test hypotheses -> isolate
          -> smallest fix -> verify -> review the diff -> commit gates

Enforcement Outside the Model

Most skill libraries are prompt text. If the model drops the instruction, nothing notices. superpowers-plus puts the checks where the model can't skip them:

• Git hooks run the commit gate chain on every git commit and git push. Changes to skills/ need a passing code-review sentinel tied to the reviewed content before the commit goes through.
• Claude Code lifecycle hooks (including SessionStart, UserPromptSubmit, PreCompact, and PreToolUse) route prompts to skills, save a resume prompt before context compaction, block pushes and branch deletions until a human approves them in the session, and stop commits made under the wrong git identity.
• CI repeats the IP scan on the PR title and body, because a squash merge builds its commit message from those fields, not from any local commit.

The AI-Harness: Instruction Context as a Budget

Every skill an agent loads stays in its context window and is re-sent on every model call. superpowers-plus treats that context as a budgeted resource and runs a closed loop over it:

Measured results so far, from the reduction ledger:

One rule outranks the byte savings: hard gates and "never" rules always stay in the kernel. That rule is enforced by review, not assumed: review caught all three of these splits moving or deleting safety content (anti-rubber-stamping rules, a required search step, a write-back procedure), and each was restored. Design: docs/harness/README.md.

Standout Skills

All 124 skills: docs/SKILLS.md. How they connect: docs/SKILL_TAXONOMY.md.

Platform Support

What's Included

Building on It

obra/superpowers (Jesse Vincent, MIT)
    └── superpowers-plus (this repo)
            └── your-org-skills (private overlay)

All 14 obra/superpowers skills are bundled here, nine of them hardened with extra enforcement gates; upstream changes are merged periodically. Teams can add a private repo on top that shadows or extends these skills with their own issue tracker, wiki, and conventions. See the Enterprise Adopters Guide, including its security checklist for overlays.

Documentation

License

MIT