Superpowers Passes 322,000 GitHub Stars as a Set of 15 Prompted Skills for 16 Coding Agents
Software / analysis
Superpowers Passes 322,000 GitHub Stars as a Set of 15 Prompted Skills for 16 Coding Agents
Jesse Vincent's MIT-licensed framework tells agents to brainstorm, plan and test first. Its clone count is large, and its evidence for better results is one author's view.
Superpowers, an MIT-licensed set of skills for coding agents, shows about 322,000 stars, 28,600 forks and 687 commits on its GitHub page. It surged again on GitHub's trending list, and it is worth asking what is actually spreading: a program, or a way of working written down as prompts. Only the second reading fits what the README describes.
The README, by Jesse Vincent and his team at Prime Radiant, calls it "a complete software development methodology for your coding agents". The code is mostly instructions an agent reads before it acts.

What the 15 skills enforce
The README groups them by purpose, and collaboration dominates.
- Collaboration9 skills
- Debugging3 skills
- Meta2 skills
- Testing1 skills
Source: github.com/obra/superpowers README, accessed 2026-11-20
The workflow runs in a fixed order: brainstorming before code, git worktrees, writing plans, subagent-driven development or plan execution, test-driven development, code review between tasks, then finishing the branch. The README says "The agent checks for relevant skills before any task" and describes these as "Mandatory workflows, not suggestions."
The friction is installation. Superpowers installs separately for each harness. Claude Code takes /plugin install superpowers@claude-plugins-official, Gemini CLI takes gemini extensions install https://github.com/obra/superpowers, and Pi takes pi install git:github.com/obra/superpowers. A team that runs three of the 16 agents repeats the step three times.
What Vincent says about the numbers
In a post on October 9, 2026, marking one year since the first public release and the launch of Superpowers 7.0, Vincent wrote that the project was approaching 300,000 stars, which he put at 12th by stars of all time. He cited more than 1.1M installs, calling it the number one third-party plugin for Claude Code.
He also gave a clone count: about 50M every two weeks from 1.95M unique IPs, and he cautioned that some of those IPs are corporate firewalls. That caveat is the useful part. Clones from continuous integration, mirrors and shared firewalls inflate the figure, so it measures traffic and not people who finished a task with it.
Stars are a vote to bookmark. The two sources also disagree on the date: the October post says approaching 300k, and the README page read this week shows about 322k, a gap consistent with continued growth but not precise.
Where it came from
Vincent wrote that Superpowers 1.0 shipped on October 9, 2025, the same day he wrote his last line of code, a single line of shell script. He says the project grew out of managing teams of junior engineers and later teams of agents. Version 6.4.2 reworked the planning process for newer models, which means the skills are tuned to model behaviour and will need retuning as models change. His own caveat is that agentic engineering does not remove obligations for safety-critical or regulated systems.
Who it helps, and who it slows down
A solo developer with a half-formed idea gets a guided start: the brainstorming skill, which Vincent says 7.0 rebuilt, is meant to help the user understand and explain what they are building, and to step aside when the user already knows. A developer who wants a one-line fix gets the same ceremony of planning, worktrees and review unless the agent judges the skill irrelevant. The README says the agent checks for relevant skills before any task, so the deciding judgement sits inside the model.
The supported list is wide: Claude Code, Antigravity, Codex App, Codex CLI, Cursor, Devin CLI, Factory Droid, Gemini CLI, GitHub Copilot CLI, Grok Build CLI, Kimi Code, OpenCode, Pi, Qwen Code, Hermes Agent and Muse. Reaching that many harnesses without per-agent forks is the project's real engineering claim. It is also the cost of the install-per-harness step above.
The case it has not yet made
Vincent's argument is that "good process management and guidance still gets a better end result" than letting a model run unsupervised. It is a plausible claim, and the post offers experience managing teams of junior engineers and later teams of agents as support. Neither source offers a benchmark, a controlled comparison or an error rate.
There is also a tension with the README itself. The contributing guidelines say maintainers generally do not accept new skills and require updates to work across all supported agents. That keeps 16 harnesses consistent, and it also limits how far the set can grow through outside contributions.
| Fact | Value | Source |
|---|---|---|
| Licence | MIT | README |
| Supported agents | 16 | README |
| Skills | 15 | README |
| Releases listed on GitHub page | None shown | README capture |
| Version tracking | .version-bump.json | README |
The last two rows are the odd ones. The page's Releases section was empty when fetched, though Vincent announced 7.0 on October 9. The README tracks versions in a file, so a user who wants to know what changed has to read the release notes his post links.
Similar agent tooling is being measured elsewhere: the Windows 11 telemetry report covers detection of Claude and Codex, and Onlook's launch reports another agent project with its star count. The number to watch is the project's issue count, 198 on the page, since a rule set that agents must follow will show its limits there first.
Sources
More in Software
- 01OpenAI's WebRTC Voice Stack Draws a 30 ms Objection That Its Own Engineers Never AddressedLuke Curley says WebRTC drops audio older than 30 ms. OpenAI's engineers wrote about port ranges and routing, which is a different problem.
- 02GitHub Disables 144 Picofly Repositories After a Nintendo Notice Names OneThe November 6 notice is the first on GitHub aimed at Switch modchip firmware, and it asked for forks to be taken down without listing a single one.
- 03Chrome's Gemini Nano Download: What Google's Docs Say, and the Numbers They Leave OutGoogle's page on built-in model management gives triggers, a 30-day rule and full-download updates, but no file size, no disk threshold and no consumer off switch.
- 04Zerodha's Blobasaur Stores Kite Layouts in Sharded SQLite, and Gives Up Durability to Do ItThe Rust, AGPLv3 blob store speaks the Redis protocol, shrank sampled production writes about 11 times, and has no live rebalancing or automatic failover yet.