ninja nerdingSuperpowers
Superpowers
15 skills. A methodology, not a suggestion. The agent checks for a skill before any task, including the first reply.
Original source
Seven signals
The whole method in short boxes. Each links to the part of this page that shows it.
- 01
Skills fire on their own
using-superpowers makes the agent look for a matching skill before any reply, even a plain question. You don't have to remember the names.
How - 02
Design before code
brainstorming asks what you really want, offers alternatives, and shows the design in chunks short enough to read. Code waits until you sign off.
How - 03
Every job gets its own workspace
After the design is approved, the work moves to a new branch in a worktree, with a clean test run before anyone writes code.
How - 04
Plans a junior could follow
writing-plans cuts the work into 2 to 5 minute tasks. Each has the file paths, the code and a way to check it.
How - 05
A fresh agent for each task
subagent-driven-development hands each task to a new subagent and reviews it before the next. executing-plans is the cheaper lane: one session, one review at the end.
How - 06
Tests first, always
Red, green, refactor. Code written before a failing test gets deleted.
How - 07
Evidence over claims
Nothing is done until it is proven. systematic-debugging finds the root cause, verification-before-completion checks the fix.
How
Words on this page
Eleven terms, in plain words. Skip this if you already work with coding agents.
- Skill
- A short instruction file the agent reads when a task matches. Each name in the catalog below is one skill.
- Plugin
- A bundle of skills you install into your coding agent in one go. Superpowers is one plugin with all of these skills inside.
- Spec
- The written design: what you are building and why. brainstorming saves it before any code.
- Branch
- A separate line of work in git, so changes don't touch the main copy until you merge them.
- Worktree
- A second folder of the same project, checked out on its own branch. Lets work happen in isolation without disturbing your main folder.
- Implementation plan
- The list of small tasks that turns the spec into code, in order, each with a check.
- Subagent
- A fresh helper agent the main agent starts for one task. It begins with a clean slate and reports back.
- TDD (red, green, refactor)
- Test-driven development. Write a test that fails (red), write just enough code to pass it (green), then tidy up (refactor).
- Code review
- Checking finished work against the plan before moving on. Superpowers rates each issue by severity; critical ones stop the next task.
- Root cause
- The real reason a bug happens, not the first symptom you see.
- PR (pull request)
- A request to merge a branch into the main code, usually so someone can review it first.
Brainstorm → worktree → plan → implement → TDD → review → finish
Official basic workflow. Brainstorm before code. Worktree after the spec. Plan before implement. TDD during the work. Review between tasks. Finish when the tasks are done.
01
brainstorming
You have an idea. No code yet.
Questions, alternatives, design in chunks you can actually read. Saves the spec.
02
using-git-worktrees
The design is approved. Isolate the work.
New branch in a worktree. Project setup. Clean test baseline before anyone writes code.
03
writing-plans
The spec is signed off. You need tasks a junior could follow.
Bite-sized steps with file paths, code, and a check for each. 2–5 minutes a task.
04
subagent-driven-development
The plan is ready. Time to implement.
Fresh subagent per task, then review. Or executing-plans: one session, one review at the end.
05
test-driven-development
A task is in progress. Tests come first.
Red, green, refactor. Deletes code written before a failing test.
06
requesting-code-review
A task just landed. Check it before the next one.
Reviews against the plan. Severity on each issue. Critical blocks the next task.
07
finishing-a-development-branch
The tasks are done. Close the loop.
Verify tests. Merge, PR, keep, or discard. Clean up the worktree.
Then back to brainstorming. The next idea starts the same loop. Skills fire again before any code.
Every skill
15 skills. One line each. IDs are what you type.
The path
Official basic workflow. In this order.
brainstorming
Questions, alternatives, design in chunks you can actually read. Saves the spec.
using-git-worktrees
New branch in a worktree. Project setup. Clean test baseline before anyone writes code.
writing-plans
Bite-sized steps with file paths, code, and a check for each. 2–5 minutes a task.
subagent-driven-development
Fresh subagent per task, then review. Or executing-plans: one session, one review at the end.
test-driven-development
Red, green, refactor. Deletes code written before a failing test.
requesting-code-review
Reviews against the plan. Severity on each issue. Critical blocks the next task.
finishing-a-development-branch
Verify tests. Merge, PR, keep, or discard. Clean up the worktree.
Also on the path
Same loop, different intensity or a side door.
executing-plans
Inline execution. One fresh review of the whole branch at the end.
dispatching-parallel-agents
One focused agent per domain. No shared state. You integrate.
receiving-code-review
Read, restate, verify against the code, then fix or push back.
Debugging
When something is wrong. Not a brainstorm.
systematic-debugging
Four-phase root cause. Trace, defend in depth, wait on conditions.
verification-before-completion
Prove the fix. Claims do not ship.
diagnosing-superpowers
Reads the transcript, cites lines, can package a scrubbed bug report.
Meta
How the system finds skills, and how you add one.
using-superpowers
Find and invoke the matching skill before a reply, including questions.
writing-skills
TDD for process docs. Fail a pressure test first, then write the skill.
Try it, smallest version first
Five steps. The install command and the test message are from the Superpowers README. The starter prompt at the end is FakeNerd's.
- 01
Install the plugin
In Claude Code, this installs it from Anthropic's official marketplace. The README has a separate install for Codex, Cursor, Gemini CLI, Copilot CLI, OpenCode, Pi and others. If you use more than one agent, install it in each.
Install for other agentsCommand
/plugin install superpowers@claude-plugins-official
- 02
Start a fresh session and test it
The README's check: send a small build request. A working install starts brainstorming before any code is written.
Command
Let's make a react todo list
- 03
Answer the questions, then approve the design
brainstorming asks, offers options and shows the design in sections. Read each one. Nothing gets built until you say it's right.
The path - 04
Say go on the plan
After the plan is written, the agent works through the tasks, with a review after each. You come back to a branch you can merge, open as a PR, keep or throw away.
The path - 05
If a session goes wrong, ask it why
Say this and the agent runs diagnosing-superpowers: it reads the transcript and shows what went wrong, line by line.
Debugging skillsCommand
figure out what went wrong with superpowers in this session
FakeNerd starter: your first real task
I want to build [one small thing, e.g. a one-page reading list where I can add and tick off books]. Before any code, brainstorm with me: one question at a time, then show me the design in short sections. When I approve it, write the plan and stop. Wait for me to say go.
FakeNerd's take
It drives, you approve
Superpowers runs the whole loop for you: the skills fire on their own and the README calls them mandatory. Its sibling, Pocock Skills, waits for you to type. Pick this one if you'd rather sign off than steer.
Same bet as Pocock: questions first
brainstorming here, grill-with-docs there. Both say most bad agent work comes from a misunderstanding, not bad code, and both fix it before a line is written.
Heavy for a one-line fix
Worktree, plan, a subagent and a review per task is a lot of ceremony for a typo. executing-plans is the lighter lane. For tiny changes, say so up front.
Proof is the best habit to steal
Even without the plugin, “show me the test run before you say done” is worth adding to any agent. That's the core of verification-before-completion.
Related on this site: Pocock Skills, the sibling set: smaller pieces you start yourself. Compound Engineering, another skills plugin with its own loop.
Attribution
- obra/superpowers by Jesse Vincent and Prime Radiant, MIT license. The skill names, the workflow order, the install command and the test message are from the repo and its README, summarized in our words.
- Prime Radiant sells commercial support for Superpowers. The optional visual companion in brainstorming loads a Prime Radiant logo that tells them which version you run; the README says to set SUPERPOWERS_DISABLE_TELEMETRY to turn it off.
- The one-line skill summaries, the signals, the glossary, the starter prompt and the takes are FakeNerd's. Not sponsored by anyone on this page.