
Coding-agent methodology maintained...
Agents that edit code immediately are often still misaligned on the requirement. This flow asks questions first, turns the idea into a design, and shows it in sections short enough to read before you approve. Only then does it write an implementation plan, with failing tests first and only the work the current task needs. Execution either gives each task a fresh subagent plus a review, or finishes in the current session and reviews once at the end. Branch wrap-up checks tests and offers merge, pull request, or discard. It fits developers who want a coding assistant to follow a fixed engineering sequence instead of skipping design and tests.
Superpowers is a coding-agent methodology maintained by Jesse Vincent and Prime Radiant. Composable skills and startup instructions make the assistant clarify the goal and confirm a design in short sections before writing an implementation plan with file paths and verification steps. After approval it can dispatch a subagent per task with review, or execute the plan in the current session. Skills trigger with the session and cover test-driven development, systematic debugging, and branch wrap-up. The code is under a permissive license. Enterprises that want commercial support can contact the team separately.
Difficulty: Intermediate
Questions before code
Brainstorming refines the idea through questions, presents the design in sections, saves a design document, and waits for approval before implementation.
Plans with verification steps
writing-plans breaks the work into small tasks. Each task names file paths, includes the code, and states how to verify it.
Subagents or one session
Each task can go to a fresh subagent with a review between tasks, or the whole plan can run in the current session with one review at the end.
Test-first work and branch finish
Implementation starts by watching a test fail, then writing code. When tasks are done, tests are checked and the choices include merge, pull request, or discard.
Align on design for a new feature
A developer describes a feature. The assistant asks questions, shows a design in sections, and writes the plan only after approval.
Dispatch tasks and review them
After the plan is approved, a subagent completes each task and the result is reviewed against the plan. Severe issues block the next step.
Find why a skill did not fire
When a session drifts, the assistant runs the diagnosis skill, reads the transcript, and reports which step ignored the skill.
No. The readme says that when the assistant sees you building something, it first asks what you are actually trying to do and shows the design in sections short enough to confirm. Only after that does it write an implementation plan, with failing tests first. Task execution starts after you say to go. It changes the order of work rather than skipping the requirements conversation.
One is subagent-driven: a fresh subagent takes each task and a review follows, first for spec compliance and then for code quality. The other executes every task in the current session and reviews the whole branch once at the end. The first is more thorough; the second costs less. Critical issues block progress. In both cases the assistant is supposed to check for a relevant skill before acting.
The repository says it is built by Jesse Vincent and the rest of Prime Radiant, with a community chat and an issue tracker. Companies that want commercial support, extra tooling, or managed spending can use the sales address on the page. The software itself is under a permissive license. Optional version telemetry includes the Superpowers version in use and not the project, prompt, or assistant, and it can be disabled with an environment variable.
Real reviews and feedback from users