Superpowers

How it works

The workflow

A session-start hook injects the using-superpowers skill, which tells the agent to check for a relevant skill before every response: "If you think there is even a 1% chance a skill might apply to what you are doing, you ABSOLUTELY MUST invoke the skill." From there the skills chain into a fixed sequence.

1. Brainstorming. The agent classifies the request as a spike, bounded work, or architectural work, and asks clarifying questions one at a time. For architectural work it proposes two or three approaches, presents the design in sections for approval, and commits a spec to docs/superpowers/specs/. Bounded work ends with a short design in the chat; a spike ends with a recommendation. No implementation starts until you approve.
2. Planning (architectural work only). writing-plans writes a plan to docs/superpowers/plans/ with global constraints and tasks, each listing files, interfaces, and test-first steps. You review the plan and choose how to run it.
3. Isolation. using-git-worktrees asks before creating a worktree, prefers the host's own worktree tool, runs project setup, and runs the tests to get a baseline.
4. Execution, in one of two modes:
   • Subagent-driven (the default): a fresh implementer subagent per task, a reviewer subagent after each task, a fix loop of up to five rounds, and a final whole-branch review on the most capable model available.
   • Inline: the session implements every task itself, and one reviewer checks the branch at the end. Cheaper, but no per-task review.
5. Discipline during execution. Test-driven development, systematic debugging when output does not match expectations, and verification before any claim that something works.
6. Finish. finishing-a-development-branch runs the full test suite, then offers three choices: merge locally, push and open a pull request, or keep the branch. Discarding work requires you to type discard.

Components

• 15 skills: using-superpowers (the router), brainstorming, writing-plans, using-git-worktrees, subagent-driven-development, executing-plans, test-driven-development, systematic-debugging, verification-before-completion, requesting-code-review, receiving-code-review, finishing-a-development-branch, dispatching-parallel-agents, writing-skills, and diagnosing-superpowers, which analyzes a session that went wrong.
• No slash commands. They were removed in version 5.1.0; skills trigger on their own or by name.
• No subagent definition files. Subagents are general-purpose agents filled in from 16 prompt templates stored inside the skills (implementer, task reviewer, re-reviewer, final code reviewer, and 11 diagnosis analysts).
• One hook: SessionStart, which injects using-superpowers. On Claude Code it fires again after /clear and after context compaction.
• Helper scripts inside skills: task-start and task-done track progress, task-brief and review-package hand work to subagents as files, and find-polluter.sh finds the test that pollutes shared state.

How it steers the agent

Everything is instruction-level. The only hook injects text; nothing blocks a tool call.

• The injected bootstrap is framed as <EXTREMELY_IMPORTANT> and says "YOU DO NOT HAVE A CHOICE. YOU MUST USE IT."
• Four skills open with an "Iron Law", for example "NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST".
• Twelve of the 15 skills contain a table of excuses the agent might make, each paired with a rebuttal.
• Checklists become todo items, and subagent-driven work keeps a ledger file, because "Conversation memory does not survive compaction."
• Reviewer subagents are told "Do Not Trust the Report" from the implementer, and the controller may not fix findings itself.
• The authors test each skill against pressure scenarios: run the scenario without the skill, watch the agent fail, then write the skill to close the gap. Behavior evals live in a separate repository, superpowers-evals.

Files it writes

• Specs: docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md, committed.
• Plans: docs/superpowers/plans/YYYY-MM-DD-<feature>.md, written to the repository; the skill does not say to commit them.
• Progress ledger, task briefs, and review packages: .superpowers/sdd/<plan>/, git-ignored and deleted once the final review is clean. A resumed session reads the ledger to skip finished tasks.

There is no memory or learnings file. Later sessions see earlier specs and plans only if the agent happens to read them while exploring the project.

Install

Each host needs its own install. On Claude Code:

/plugin install superpowers@claude-plugins-official

Or from the author's marketplace: /plugin marketplace add obra/superpowers-marketplace, then /plugin install superpowers@superpowers-marketplace.

Other hosts have their own commands in the README, for example /add-plugin superpowers in Cursor, /plugins in Codex, pi install git:github.com/obra/superpowers in Pi, and gemini extensions install https://github.com/obra/superpowers in Gemini CLI. Gemini CLI support was removed in 6.1.0 and restored in 6.2.0 while its future is evaluated. Codex needs multi_agent = true in its config for subagents.

Needs git and bash (Git Bash on Windows). Node.js is needed only for the optional visual companion.

When to choose it

• You want one fixed process for every task, from questions through spec, plan, tests, review, and merge, without anyone having to type commands.
• Your team uses one of the 15 supported hosts, or several of them.
• You care more about catching problems before merge than about tokens per task.

When not to

• You want the agent to remember lessons between sessions. Superpowers has no learnings store.
• You need hard gates. Every rule is text the model can, in principle, ignore.
• You want a stable process. The project published 45 releases between October 2025 and September 2026, several of which rewrote skills.
• You want a small footprint per task. Subagent-driven mode spends a fresh context on every task and every review.

Parts of the README describe older behavior, such as "two-stage review" and "2-5 minute" tasks. The skills themselves are current; trust them over the README.

Sources

• obra/superpowers at 8ca22db: README.md, RELEASE-NOTES.md, AGENTS.md, skills/*/SKILL.md, hooks/, and the plugin manifests
• superpowers-evals (linked, not checked)
• DeepWiki was not used: it was unreachable from the research environment.