Image: GitHub
Why it mattersA coding agent that asks what you want before writing code wastes fewer runs. Superpowers plugs that step into Claude Code, Codex and ten other agents.
Most coding agents open a file and write code. The project called Superpowers makes them stop and ask what you are trying to do. The GitHub repository has 308,863 stars under an MIT licence. It sits 24th on Trendshift's daily list of trending projects.
The author is Jesse Vincent, who publishes on GitHub as obra. His team at Prime Radiant maintains the project. The repository was first opened on 9 October 2025. The star count is a total, not a weekly gain, because the repository is just over a year old.
What the framework does
The README calls Superpowers "a complete software development methodology for your coding agents". Every task runs through five phases. Discovery comes first. The agent reads the project files and asks what the user really wants. Design comes next. The agent writes a short design document and the user approves it. Planning turns the design into a task list. Implementation runs each task. Review happens between stages, with a dedicated skill called requesting-code-review.
The twelve named skills group into three families. Testing and debugging covers test-driven-development, systematic-debugging and verification-before-completion. Collaboration covers brainstorming, writing-plans, subagent-driven-development and using-git-worktrees. Meta covers writing-skills and using-superpowers. Each one is a Markdown file the agent reads at the right moment.
Which agents it works with
The README lists thirteen compatible coding agents. The list covers Claude Code, Cursor, Codex, Devin CLI, Gemini CLI and GitHub Copilot CLI. It also covers Grok Build CLI, Kimi Code, OpenCode, Pi, Qwen Code, Hermes Agent and Muse. Each agent has its own install command. The author also lists Antigravity, which was added in v7.0.0 on 10 October.
For a reader with Claude Code already installed, the framework works through Claude Code's skills loader. For a reader on Codex or Cursor, it works as plain Markdown files the agent reads from a known location.
What changed in the latest release
Version 7.0.1 was published on 21 October. The release notes say the brainstorming skill now reads existing project files before it asks questions. The author gives one example from a user session before this change. The fix loop ran 18 rounds instead of a planned 5. Three skills were restructured to stay under a 20,000-character limit set by Claude Code: subagent-driven-development, writing-skills and executing-plans.
The v7.0.0 release on 10 October rebuilt the brainstorming step from scratch. The author says the new version "draws out what you want before it proposes anything". The older version opened every brainstorm with the same generic questions.
What it costs the user
Everything runs in the user's own agent. There is no hosted service and no sign-in. The user pays for the model calls their coding agent already pays for. The install is a shell command for each agent they use.
The author says a human still approves the design before the agent writes code. The framework is a methodology, not an autopilot. For a reader who runs their coding agent on every change, the cost is five extra minutes a run. The gain is fewer retries when the agent misreads the task.
Source
obra/superpowers, on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.