What is Superpowers? Where it is a Must, and Where It Is Not

The Superpowers plugin for Claude Code, a set of 15 skills whose GitHub repository has about 294,000 stars, makes coding agents ask clarifying questions and present a plan before writing code, according to the plugin's README and release notes. Only one skill, using-superpowers (about 490 words), loads at session start, while the other 14 stay on disk; the plugin also lists support for Codex, Cursor, Copilot CLI and Gemini CLI. Release notes cite a session where "that scope is ok" was taken as permission to start building and 11 of 12 test runs where the agent ran only the one test file it was told about, prompting changes in versions 6.4.1 and 6.4.2.

Hi everyone. You ask a coding agent: “Add a CSV export for the report.” It asks nothing, creates three files, and answers “Done.” You run the project. The export works, but the page next to it no longer opens. A plugin called Superpowers is meant to prevent this: it makes the agent ask first, show a plan, and only then write code. Its repository has about 294,000 stars on GitHub. A note on how this article was made: I used Claude to help read the repository and its release notes and to draft the text. I checked the numbers against the sources linked at the end. The short answer: Superpowers is a set of 15 skills for Claude Code. A skill is a text file with instructions, which I covered in an earlier article https://medium.com/towards-artificial-intelligence/what-are-skills-and-how-do-you-use-them-cursor-claude-code-e7451974bb66 . The agent gets no new abilities. Think of a preflight checklist. A pilot already knows how to fly, but even good pilots skip steps under pressure. Agents skip the same kind of steps: they start writing code before they understand the task, and they say “done” without running anything. Superpowers is that checklist, written for an agent. Only one skill, using-superpowers about 490 words , is loaded when a session starts. The other 14 stay on disk until the task calls for them, so they do not eat your context in advance. The plugin also works outside Claude Code: the README lists Codex, Cursor, Copilot CLI, Gemini CLI and others. Once you give a task, the agent works through four steps: In the CSV example, the agent would first name the type of task out loud, so you can overrule it. Then it asks one question about the goal, for instance whether the export should contain only the filtered rows or all of them, and restates how it understood you, marking its own guesses. The plan carries a section called Review Focus: up to five places the spec implies but no test covers, such as an export of a million rows. I would read the plan starting there. This is the plugin’s described behavior, not a record of my own run. The release notes explain the reasons. In one session, “that scope is ok” was taken as permission to start building, so an approval now covers only what you were actually shown. In 11 of 12 test runs, the agent ran only the one test file it was told about, and a broken test next door went unnoticed. Now “done” means the full run. Nothing technically stops the agent from skipping a step. The skills work by instruction, and the agent decides whether to follow them. Newer versions also scale the process to the task: a question gets a short experiment, a small change gets a short design in chat, and the full cycle is kept for bigger work. The plan is the one document worth reading before work starts. Begin with Review Focus. Then compare the plan with the spec: a plan several times longer has probably turned into a retelling of the code, and the authors added this check themselves in version 6.4.2. A good plan names files, signatures and what each test checks, and it does not need whole functions. After approval, there are two modes. In Native mode the agent does all the tasks itself and a fresh agent reviews the whole branch at the end. It is the cheapest. In Subagent-driven mode every task goes to a new agent with a review after each one, which is more thorough and more expensive. Native used to be a 64-line stub that, by the maintainers’ own measurement, performed the same as no plugin at all, and they rebuilt it in version 6.4.1. Native when the plan is clear, Subagent-driven when the tasks are independent or the risk is high. The best independent test I found was run by Norbert Laszlo https://norbert-laszlo.medium.com/can-a-plugin-improve-codex-benchmarking-the-superpowers-plugin-05d020066565 in June 2026: 500 tasks on Codex, with and without the plugin. Three caveats apply. It was Codex, not Claude Code. The tasks were mostly bug fixes. And the test ran before version 6.0. I have not seen an independent test of the current versions, and I would like to see one. Cached tokens cost less than fresh ones, so 40% more tokens does not mean 40% more money. On a flat-rate plan you mostly pay in time. the process is worth it when a mistake costs more than the process. It is a must for: It is not needed for: “A must” is my own rule here. There is no independent measurement for new features, only the authors’ own log of failures, so test it on your tasks. A team is a separate case: a named process is easier to discuss than everyone’s personal prompts, and the vanja.io author lists that as one argument for adopting one. In my own work I reach for it on epics, like redesigning a screen from a Figma or Claude Design mockup, or reverse engineering a feature, where the task has to be split into parts before anyone writes code. The best answer to “do I need it” is your own measurement: Write it down as you go, not afterwards. /plugin install superpowers@claude-plugins-official Then check the version: ls ~/.claude/plugins/cache/ /superpowers/ Mine showed 5.1.0 while the latest was 6.4.2. Plugins do not always update by themselves, and the newer versions changed a lot. Your instructions in CLAUDE.md outrank the skills, so you can switch the process off for small things. Three lines that help even without the plugin: - Edits under ~20 lines: no plans, no questions, just do it.- "Done" means: the full test command was run and its output read.- After three failed fixes, stop and rethink the approach. The optional visual companion in brainstorming loads a logo from the authors’ site, and that request includes the Superpowers version. Set SUPERPOWERS DISABLE TELEMETRY to any true value to turn it off. Have you measured what it costs on your own tasks? I’m a Product Manager with an engineering background ex SWE , focused on building, scaling and growing products. I’m especially interested in new technologies, particularly AI/ML, and how they can be applied in real workflows. Email: akzhankalimatov@gmail.com ✅ Connect with me on LinkedIn , X Twitter ✅ Press and hold the 👏 to give up to 50 claps for this article What is Superpowers? Where it is a Must, and Where It Is Not https://pub.towardsai.net/what-is-superpowers-where-it-is-a-must-and-where-it-is-not-8c540af07ffa was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.