The Superpowers Skills System
In one line: skills provide reusable process instructions; effective gates and accountable decisions provide enforcement.
Superpowers supplies composable development skills. Installation and invocation differ by host and release; consult its maintained instructions, approve the dependency and verify the installed skills/tools. Do not assume one global installation, one slash-command syntax or a fixed skill count across environments.
The reference workflow uses these capabilities:
| Capability | Example skills | Evidence of use |
|---|---|---|
| Explore and plan | brainstorming, writing-plans | Reviewed design and actionable scope |
| Execute permitted work | executing-plans, subagent-driven-development, dispatching-parallel-agents | Owned tasks, authorized effects and scoped results |
| Develop and diagnose | test-driven-development, systematic-debugging | Reproduced failures and verified behaviour |
| Verify and review | verification-before-completion, requesting-code-review, receiving-code-review | Current evidence and adjudicated findings |
| Isolate and integrate | using-git-worktrees, finishing-a-development-branch | Preserved work, completed checks and authorized integration |
| Maintain instructions | using-superpowers, writing-skills | Relevant loading and scenario-validated guidance |
These are reference mappings, not promises about every upstream release. Read the installed skill before relying on its workflow. S4U's project skills add domain/profile guidance; they do not grant permission to spawn agents, install services, merge, upload code or deploy.
A skill may guide a model to stop for missing evidence. It can still be omitted or misapplied. The shipped Stop hook is advisory and does not collect test receipts; it cannot guarantee completion discipline. Likewise, a finishing skill does not prove memory was updated or that its content remains correct.
Evaluate skills with realistic baseline and forward scenarios, including pressure to skip checks, ambiguous ownership and untrusted input. Record observed behaviour and limits. Enforce critical transitions at verified action boundaries outside the model where required. See §§3, 5 and 7 for the surrounding lifecycle and control architecture.