Fifteen skills that give a coding agent a full development process

What problem it actually solves

Superpowers is not a pile of prompt templates. It is a complete software development methodology for coding agents: composable skills plus initial instructions that stop the agent from jumping straight into code. Current version v6.4.1 (2026-09-18), MIT, 289,160 stars / 25,872 forks — the largest single skill repository we have found, and its growth comes from the methodology rather than a toy demo.

The premise is plain: models are not short on coding ability, they are short on process discipline. Without it an agent skips requirement clarification, skips the failing test, skips code review, and ships something that merely looks correct. Superpowers nails that pipeline down with 15 skills.

The 15 skills and the trunk workflow

Phase	Skill	What it forces
1. Clarify	brainstorming	Turns a vague ask into a spec, shown in chunks short enough that you actually read them
2. Isolate	using-git-worktrees	One worktree per task; side effects cannot leak into the main tree
3. Plan	writing-plans	Writes a plan an enthusiastic junior engineer with poor taste, no judgement, no context and an aversion to testing can follow
4. Execute	subagent-driven-development / executing-plans	Subagents take each task and review each other; executing-plans is the v6.4.1 rewritten Native inline mode with a single final review
5. Test	test-driven-development	Red/green TDD. Iron law: NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST — wrote the implementation first? Delete it
6. Review	requesting-code-review / receiving-code-review	Ask for review, and answer review findings with evidence
7. Close	finishing-a-development-branch / verification-before-completion	Verify before you claim completion

Also shipped: systematic-debugging (method over shotgunning), dispatching-parallel-agents, writing-skills (a skill for writing skills), and — new in v6.4.1 — diagnosing-superpowers, a session-forensics skill that requires path:line evidence for every claim.

The single best idea to steal: "Rulings, not stalls"

Most agent frameworks fail by stopping to ask the human on every uncertainty, which makes the human the bottleneck. Superpowers inverts it: rule for yourself by default, and stop only in 4 cases.

1. The action is irreversible or destructive
2. It touches a security-sensitive surface
3. The side effect lands outside the worktree
4. The plan itself is too broken to continue

Those four are portable to any homegrown agent. We adopted them in this site's own ingestion pipeline (confirm before irreversible or production writes).

Equally important: using-superpowers requires checking for an applicable skill before any response, with a threshold as low as "even a 1% chance". Skills that sit installed but unused almost always lack this trigger instruction.

Installation and harness coverage

It is cross-harness, with 16+ documented targets: Claude Code, Antigravity, Codex App/CLI, Cursor, Devin CLI, Factory Droid, Gemini CLI, GitHub Copilot CLI, Grok Build CLI, Kimi Code, OpenCode, Pi, Qwen Code, Hermes Agent, Muse. Install it separately for each harness you use.

# Claude Code (official plugin marketplace)
/plugin install superpowers@claude-plugins-official

When not to use it

It is heavy. A full pass goes spec → plan → worktree → subagents → review, which is waste for a one-line config change. It pays off on multi-file refactors, greenfield features, and team work that needs an auditable trail. subagent-driven-development also burns markedly more tokens than inline mode; v6.4.1's rewrite of executing-plans down to a single final review is a direct answer to that cost.

Its stated philosophy has four lines: TDD, systematic over ad-hoc, complexity reduction, and evidence over claims. The last one is the soul of the repo, and it is our first test for whether any agent skill deserves a place in this index.