[GitHub] [README] [Release notes] [Issues] [Upstream]
Superpowers for Google Antigravity
for the 2.0 app, the IDE and the agy CLI
Abstract
Superpowers is a set of skills that makes your Antigravity agent work like a careful engineer. It asks what you want and why, writes a design you approve, plans the work, builds it test-first with subagents, reviews every task and the whole branch, then finishes the branch. The skills trigger on their own.
1. Installation
agy plugin install https://github.com/ roundpilot/ superpowers-antigravity
git clone https://github.com/ roundpilot/ superpowers-antigravity ~/ .gemini/ config/ plugins/ superpowers
git clone https://github.com/ roundpilot/ superpowers-antigravity "$env:USERPROFILE\ .gemini\ config\ plugins\ superpowers"
Then say "Let's make a react todo list". The agent SHOULD start asking design questions instead of writing code. Workspace installs, WSL, ZIP downloads and updates: README, Installation.
Status of This Memo
This document describes a working plugin, version 7.0.2, released in November 2026 under the MIT License. Its test suite passes, and its live tests passed on Antigravity CLI 1.5.0; transcript tests and app smoke tests have not run. Distribution of this memo is unlimited.
Table of Contents
- 1. Installation1
- 2. Requirements Language2
- 3. The Workflow2
- 4. Skills3
- 5. Bundled Agents4
- 6. Differences from Upstream4
- 7. Security Considerations5
- 8. IANA Considerations5
- 9. References5
- Acknowledgements5
- Author's Address5
2. Requirements Language
The key words "MUST", "MUST NOT", "SHOULD" and "MAY" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174].
They are also how the skills talk to your agent. Superpowers is written in this register, and the agent treats a skill as a rule it follows, not a suggestion.
3. The Workflow
Each stage is a skill. Select a stage in Figure 1 to read what it requires.
what and why tasks, files test first spec review code review whole branch merge, PR, keep
3.1. Brainstorm
The agent MUST ask what you want, and why, before proposing anything. It plays its understanding back, offers two or three approaches and sizes the work. A quick, clear task just gets done. Anything bigger MUST NOT be built until you approve the design.
brainstorming
3.2. Plan
The design is broken into small tasks with exact file paths and verification steps, clear enough for an enthusiastic junior engineer with no context. You review the plan before anything runs, and work moves to its own branch.
writing-plans, using-git-worktrees
3.3. Build
A fresh subagent takes each task test-first: a failing test, the least code that passes it, a commit. A spec reviewer checks the work against the task, then a code reviewer checks its quality. Agents can run for a couple of hours without leaving the plan. The IDE has no subagents, so there the tasks run inline.
subagent-driven-development, test-driven-development, executing-plans
3.4. Final review
A read-only reviewer at the Pro tier reads the whole branch before anything merges, and reports issues by severity.
requesting-code-review, receiving-code-review
3.5. Finish
The tests are run once more, then you choose: merge, open a pull request, or keep the branch. Nothing is called done without evidence.
finishing-a-development-branch, verification-before-completion
4. Skills
Sixteen skills. Each one loads only when it applies; a single short rule stays loaded so the agent knows they exist.
4.1. Building
- brainstorming
- Draws out what you want and why, offers approaches, sizes the work and writes the design.
- writing-plans
- Turns the design into small tasks with exact files, tests and a review gate.
- subagent-driven-development
- One fresh subagent per task, each followed by a spec review and a code review.
- executing-plans
- Runs the plan inline in one session, with one review of the whole branch.
- dispatching-parallel-agents
- Runs independent work at the same time on separate branches.
- using-git-worktrees
- Isolates the work on its own branch.
- finishing-a-development-branch
- Checks the tests, then offers merge, pull request or keep.
4.2. Testing and debugging
- test-driven-development
- Red, green, refactor: the test fails first, then the least code passes it.
- systematic-debugging
- Finds the root cause in four phases before any fix.
- verification-before-completion
- No "it works" without the command output that proves it.
- browser-testing
- Screenshots and page checks as evidence for UI work.
- diagnosing-superpowers
- Works out what went wrong in a session, with line-level evidence.
4.3. Review and meta
- requesting-code-review
- Sends a task or the whole branch to a read-only reviewer.
- receiving-code-review
- Weighs feedback on its merits before changing anything.
- writing-skills
- Writes and tests new skills.
- using-superpowers
- The short rule that tells the agent the skills exist.
5. Bundled Agents
Subagents ship with the plugin. "inherit" runs your session's model; "flash" and "pro" are Antigravity's own tiers, and which model each runs depends on your plan.
| Agent | Tier | Role |
|---|---|---|
| implementer | inherit | Builds one plan task and fixes review findings |
| implementer-flash | flash | The same, for fully specified tasks |
| implementer-pro | pro | The same, for escalations and design work |
| spec-reviewer | inherit | Read-only: checks a task against its text |
| code-reviewer | inherit | Read-only: reviews one task or a small range |
| re-reviewer | inherit | Read-only: verifies a fix round |
| final-reviewer | pro | Read-only: reviews the whole branch |
| session-analyst | inherit | Read-only: reads transcripts for diagnosis |
| bundle-scrubber | inherit | Redacts a diagnosis bundle |
6. Differences from Upstream
This is a fork of obra/superpowers [SUPERPOWERS], rebuilt for Antigravity only.
- Skills name Antigravity's own tools, such as view_file, run_command and invoke_subagent, with no translation layer.
- Reviewers are bundled read-only agents; implementers come at three tiers.
- An always-on rule loads the bootstrap in the 2.0 app, the IDE and the CLI.
- In the IDE, which has no subagents, plans run inline.
For Claude Code, Codex, Cursor and other agents, you SHOULD use upstream instead.
7. Security Considerations
Skills and agents are plain Markdown. The plugin ships no MCP servers and installs no hooks; an optional check script is for contributors only. Reviewer agents are read-only. Work happens on its own branch, and nothing merges until you choose to.
8. IANA Considerations
This document has no IANA actions.
9. References
9.1. Normative References
- SUPERPOWERS
- Vincent, J., "Superpowers", https://github.com/obra/superpowers.
- ANTIGRAVITY
- Google, "Antigravity documentation", https://antigravity.google/docs.
- RFC2119
- Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997.
- RFC8174
- Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, May 2017.
9.2. Informative References
- README
- "Superpowers for Google Antigravity", README.md.
- NOTES
- "Release notes", RELEASE-NOTES.md.
Acknowledgements
Superpowers was created by Jesse Vincent. This port keeps his methodology and rebuilds it for Antigravity. If it helps you, consider sponsoring the original project.
Author's Address
roundpilot
GitHub: https://github.com/roundpilot
Issues: superpowers-antigravity/issues
Request for Install
agy plugin install https://github.com/ roundpilot/ superpowers-antigravity
2.0 app and IDE: Section 1. MIT License.