Superpowers: Agentic Development Framework
/superpowers
Superpowers is a skills framework that forces coding agents through a brainstorm-plan-TDD-review pipeline instead of letting them jump straight to code. Reach for it if you run Claude Code, Cursor, Codex, or one of the other eight supported harnesses and want that pipeline enforced automatically on every task. Skip it if your agent's harness isn't on that list, or you'd rather keep the process loose.
What Superpowers Does for Coding Agents
Superpowers is an agentic skills framework: a library of composable skills plus a bootstrap instruction that makes coding agents use them without being asked. It walks an agent from brainstorming a spec, through a written implementation plan, into RED-GREEN-REFACTOR TDD and a two-stage code review, before anything merges. Eleven agent harnesses can install it, including Claude Code, Cursor, and Codex.
Core Skills and Workflow Automation
- ✓Skills trigger automatically once installed — no slash command or manual invocation needed to start the pipeline.
- ✓The test-driven-development skill enforces RED-GREEN-REFACTOR and deletes any code written before its test existed, not just a reminder to write tests.
- ✓subagent-driven-development dispatches a fresh subagent per task with two-stage review — spec compliance first, then code quality — so one agent isn't grading its own work.
- ✓systematic-debugging runs a 4-phase root-cause process, including root-cause-tracing and defense-in-depth techniques, instead of guess-and-check patching.
- ✓using-git-worktrees creates an isolated branch and verifies a clean test baseline before any implementation code gets written.
- ✓writing-skills is a meta-skill for authoring new skills in the same format, though the maintainers say they generally don't take outside skill contributions.
The Superpowers Development Workflow
The workflow runs as seven skills in sequence. brainstorming activates first, asking questions and showing the design in readable chunks before you sign off. using-git-worktrees then creates an isolated branch and confirms a clean test baseline. writing-plans breaks the approved design into 2-5 minute tasks, each with exact file paths and complete code. Once you say "go," either subagent-driven-development (fresh subagent per task, two-stage review) or executing-plans (batch execution with human checkpoints) does the implementation work under test-driven-development's RED-GREEN-REFACTOR discipline. requesting-code-review checks each task against the plan and blocks progress on critical issues. finishing-a-development-branch runs last, verifying tests and presenting merge, PR, keep, or discard options.
Installing Superpowers for Your Agent
Installation is documented and differs by harness — if you use more than one agent, you install Superpowers separately in each. In Claude Code, run `/plugin install superpowers@claude-plugins-official` from the official marketplace, or register `/plugin marketplace add obra/superpowers-marketplace` first for the community marketplace variant. Cursor uses `/add-plugin superpowers` in Agent chat. Gemini CLI uses `gemini extensions install https://github.com/obra/superpowers`. Codex CLI, Codex App, Factory Droid, GitHub Copilot CLI, and Kimi Code each install through their own plugin marketplace command. Pi installs via `pi install git:github.com/obra/superpowers`, or loads a local checkout with `pi -e /path/to/superpowers` for development. OpenCode requires fetching install instructions from a raw GitHub URL inside the agent itself.
Strengths
- ✓The workflow is mandatory rather than advisory: skills trigger automatically, so the design and test steps aren't easy to skip mid-task the way a written guideline is.
- ✓TDD enforcement is mechanical — the test-driven-development skill deletes code written before its test, not just a style note.
- ✓Two-stage subagent review (spec compliance, then code quality) catches issues per task instead of after a multi-hour autonomous run has already compounded them.
- ✓Covers eleven different agent harnesses with one methodology, so a team split across Claude Code, Cursor, and Codex can standardize on the same process.
- ✓MIT-licensed and free, with an optional paid line to Prime Radiant for enterprise support and managed spending.
Considerations for Using Superpowers
- △Requires a separate install and update for every harness you use — there's no single global install that covers Claude Code, Cursor, and the rest at once.
- △The project generally doesn't accept outside contributions of new skills, and any accepted skill update has to keep working across all eleven supported agents, which narrows what community changes look like.
- △It's not a standalone tool — it depends entirely on one of the eleven supported agent harnesses to run at all.
- △The brainstorming skill's optional visual companion loads a Prime Radiant logo from their website by default, sending the Superpowers version in use; it's opt-out via the SUPERPOWERS_DISABLE_TELEMETRY environment variable, not opt-in.
- △Updating is, per the README, "somewhat coding-agent dependent" — not a single guaranteed command across every harness.
Similar AI Agent Frameworks
Common Questions About Superpowers
Superpowers installs into eleven harnesses: Claude Code, Antigravity, Codex App, Codex CLI, Cursor, Factory Droid, Gemini CLI, GitHub Copilot CLI, Kimi Code, OpenCode, and Pi, each with its own install step.
The test-driven-development skill runs RED-GREEN-REFACTOR: write a failing test, watch it fail, write minimal code to pass, refactor, commit. Code written before its test gets deleted.
Superpowers generally doesn't accept outside contributions of new skills, per its README, since any skill update has to keep working across every supported coding agent.
Superpowers doesn't collect project, prompt, or click data. Its optional visual companion loads a logo from Prime Radiant's site by default and sends only the Superpowers version, disabled via SUPERPOWERS_DISABLE_TELEMETRY.
Superpowers is MIT-licensed and free for enterprise use. Prime Radiant, the team behind it, also offers paid commercial support and managed spending on request at [email protected].
Superpowers is released under the MIT License, per the LICENSE file in the obra/superpowers GitHub repository.
The problem it solves
Coding agents left to their own devices tend to skip straight to writing code, skim past tests, and lose the thread on long autonomous runs with no checkpoint forcing a review. Superpowers targets that specific failure mode: an agent working unattended for a couple of hours needs brainstorming, planning, TDD, and code-review gates baked into the harness itself, not a hope that a system prompt covers it.
Best use cases
- •Long autonomous coding sessions: subagent-driven-development dispatches a fresh subagent per 2-5 minute task with two-stage review, which is what the README describes as letting an agent work for a couple of hours without deviating from the plan.
- •Teams standardizing TDD across different AI coding tools who want the same RED-GREEN-REFACTOR discipline enforced instead of relying on each agent's own judgment.
- •Debugging agent-introduced regressions with the systematic-debugging skill's 4-phase process instead of ad hoc guessing.
- •Solo developers who want a forced design step before code, via brainstorming's chunked spec review.
Who should try it — and who should skip
Try Superpowers if you already run Claude Code, Cursor, Codex, Gemini CLI, or one of the other supported harnesses and want brainstorming, planning, TDD, and review enforced instead of optional. Skip it if your harness isn't on the list, you don't want a per-harness install and update routine, or you want to submit custom skills upstream — the project generally doesn't take them.
Related repositories
Still deciding about superpowers?
One click hands the question to an AI along with this page — see what it says about superpowers.