Superpowers Claude skills

Installing Superpowers from GitHub on each agent

The README states that installation differs by harness. If you use more than one coding agent, install Superpowers separately in each one. The FAQ above covers the official Claude Code marketplace and Codex CLI. Here are the other routes the README documents.

Claude Code, via the Superpowers marketplace. The obra/superpowers-marketplace repo is a curated Claude Code plugin marketplace. It carries Superpowers and some related plugins. Register it, then install from it:

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

Antigravity. Install directly from the GitHub repo:

agy plugin install https://github.com/obra/superpowers

To update, run the same command again. Antigravity lists each installed skill's description. The using-superpowers description tells the model to load that skill at the start of a conversation, and that is how the rest of the pack gets activated.

Codex App. Open Plugins in the sidebar, find Superpowers in the Coding section, click +, and follow the prompts.

Cursor. In Agent chat, run /add-plugin superpowers, or search for "superpowers" in the plugin marketplace.

Devin CLI. Install with devin plugins install obra/superpowers and update with devin plugins update superpowers.

Factory Droid. Register the repo as a marketplace with droid plugin marketplace add https://github.com/obra/superpowers, then run droid plugin install superpowers@superpowers.

Gemini CLI. Install the extension with gemini extensions install https://github.com/obra/superpowers.

OpenCode users coming from an older setup should read .opencode/INSTALL.md in the repo. It explains how to remove the old symlinks (~/.config/opencode/plugins/superpowers.js and ~/.config/opencode/skills/superpowers), optionally delete the cloned repo, and drop any skills.paths entry you added to opencode.json for Superpowers.

What a session looks like in practice

Superpowers does not add commands you have to remember. The skills trigger on their own. What changes is how the agent responds to a request.

1. Small, fully specified tasks go straight through. The brainstorming skill first checks whether the request already contains everything that matters. Its own example is "make the icon cornflower blue". In that case the agent makes the change and says what it did.
2. The agent reads before it asks. In an existing project, or when you point it at a spec, plan, handoff, or files, it reads that material and the code it touches. It keeps reading until it can explain how the relevant parts work today. The skill tells it never to ask something the project already answers.
3. A short first reply. If you already said what you want, it aims for about 200 words. It restates your request in its own words (what, for whom, and why) and adds anything it found that could change the ask.
4. Approaches, then a design doc. It lays out the options in a message of their own, waits for you to pick one, writes a short or full design doc, and asks you to approve it. If you want to skip the questions, you can opt out. A spike you agree to also does not count as implementation.
5. The plan. Once the design is approved, writing-plans produces an implementation plan detailed enough that, in the README's words, "an enthusiastic junior engineer with poor taste, no judgement, no project context, and an aversion to testing" could follow it. The plan stresses red/green TDD, YAGNI, and DRY.
6. Execution. When you say "go", subagent-driven development begins. Subagents take each engineering task, the main agent reviews their work against the plan, and the run continues.

This is the main difference from using Claude Code without the pack. Without it, the agent starts writing code right away. With it, you spend a few turns on intent and design before any code changes.

Supporting skills worth knowing about

Several skills sit outside the main brainstorm, plan, and execute loop.

using-superpowers is the bootstrap. It is a small SKILL.md with seven reference files, and it makes the agent use the rest of the pack.

dispatching-parallel-agents applies when you have two or more independent tasks with no shared state, for example three test files failing for different reasons. The agent sends one subagent per problem domain. Each subagent gets a specific scope, a clear goal, constraints such as "don't change other code", and an expected output. Subagents receive only the context they need, not the full session history. The skill says not to use this approach when the failures are related or when agents would interfere with each other.

writing-skills is for authoring your own skills. It is one of the larger files in the pack, with eight reference docs and a script.

diagnosing-superpowers is what you reach for when a session goes wrong: a skill fires when it shouldn't or stays silent when it should, the agent ignores its plan or repeats work, or token use is higher than expected. It asks you one question at a time until it has a clear problem statement, then finds the session transcripts on disk and creates a workspace under ~/.superpowers/diagnosing-superpowers/<session-id>/. Next it sends analyst subagents to check separate dimensions: skill timeline, plan adherence, repeated work, stumbles, quality evidence, request conflicts, and cost and time. Each finding has to cite a path:line in the transcript, and the skill discards any finding that doesn't. The output is a report you can also use as a bug report for the maintainers.