Decide whether you want the workflow
Superpowers gives coding agents a prescribed development process. The useful question is whether you want that process in your work. Our recorded sandbox run did not establish that it produces better code.
The README describes clarifying intent, agreeing on a design, creating an isolated worktree, planning, implementing through subagents or an inline executor, following test-driven development, reviewing, and finishing the branch. It calls relevant workflows mandatory. Treat this as the project's intended behavior; we did not observe an agent complete that sequence.
Task size alone is a poor reason to reject it. The current README says small changes get agreed in chat, while projects get a written design document. Teams that dislike test-first implementation still face a substantive mismatch: the documented TDD workflow says code written before tests is deleted.
What our recorded run establishes
Our run began on October 1, 2026 at 03:29:57.001 UTC and used commit 8ca22dba9a94f28898bbce59f2537ff4d87c747d, associated with v6.4.2. The recorded sandbox configuration was 3 CPUs, 8 GB of memory, and image lab-node:22. The image label is not independent runtime verification.
The setup wrapper returned exit code 0 after 6,191 ms. Its output contains an npm ci error about the required lockfile, followed by an up-to-date npm result. Calling this a six-second Superpowers installation would conceal the distinction between a repository setup command and installing a plugin in a coding agent. The record does not identify every command behind that final result.
The stored metrics report 0 dependencies, a 4 MB dependency/install footprint, and a 2 MB repository containing 229 files and about 13,156 lines of source. These are recorded scan values, not a measurement of plugin activation or development performance. The complete harness run took 22 seconds.
The npm audit summary records 0 advisories. The number of packages scanned is unknown, and the dependency count does not establish audit coverage. This result cannot support a claim that Superpowers is safe.
Skipped tests do not mean absent tests
The harness skipped build and test steps because it detected no script. It also recorded a tests directory, no Makefile, no detected CI workflows, and no Dockerfile or Compose configuration. Those findings describe the detector's scope, not an exhaustive inventory of the project's automation.
The README explicitly describes skill-behavior tests using the drill evaluation harness from superpowers-evals and plugin-infrastructure tests under tests/, run through the relevant run-*.sh or npm test. Our harness did not execute those paths. There is no passing-test count, failed product test, or measured skill-behavior result to report from this run.
No coding-agent installation, installed-binary check, session bootstrap check, or representative coding workload was recorded. Those are the next checks needed for an adoption decision.
Check the host and version separately
Installation differs by coding-agent host. The README says to install Superpowers separately in each one, so success in one host would not verify another.
Antigravity issue 2247 is historical compatibility evidence, not a current open blocker. The September 2, 2026 report concerned Superpowers 6.3.0 and Antigravity CLI 1.1.24. A later comment reported the problem on v6.4.2. On October 21, 2026, a collaborator closed the issue as fixed in v7.0.0, describing a native plugin manifest and skill-discovery bootstrap instead of the earlier hook route. The current README likewise describes skill discovery. These are upstream reports; our v6.4.2 sandbox run did not test Antigravity or v7.0.0.
Other documented caveats remain relevant to a trial. Hermes can lose the initial bootstrap after a long session compacts. Muse requires active sessions to restart after installation, with hook approval on first install. Check the instructions for your host rather than assuming every integration loads skills the same way.
Read planning claims at the right level
The v6.4.2 release notes say plans should record interfaces, test assertions, and specification values rather than routinely writing complete implementation bodies. They also replace estimated task durations with one action that has a checkable result. The README's basic-workflow summary still mentions complete code and tasks of 2-5 minutes, so those summaries do not fully agree.
For this version, the release notes give the more specific account of the planning change. Their timing, token, and planted-defect results are upstream measurements, not results from our run. They do not establish savings for your project.
Alternatives and a useful trial
Anthropic Skills is relevant when you want examples of individual Claude capabilities or a reference for writing skills. Its own disclaimer says the examples are for demonstration and education and should be tested in your environment.
OpenAI Skills is now deprecated. Its notice directs readers to OpenAI Plugins for current Codex examples and the Build plugins guide for adding skills. It remains relevant to older instructions, but recommending it as a current catalog without that warning would mislead readers.
For a Superpowers trial, install it through your host's documented route, start a fresh session where required, and check that the expected skill loads. Then use a representative change to assess whether clarification, planning, test-first work, and review help you. Record the host and version alongside the result.
The optional visual companion loads a Prime Radiant logo from the project's website and includes the Superpowers version in that request, according to the README. It documents SUPERPOWERS_DISABLE_TELEMETRY, DISABLE_TELEMETRY, and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC as opt-outs. Configure one before a trial if your policy requires it.
How to read the scores
These editorial ratings assess documented adoption readiness, not benchmarked effectiveness. The scale is 1 for substantial documented obstacles, 2 for limited supporting evidence, 3 for a documented path with important verification gaps, 4 for a well-supported path with bounded caveats, and 5 for comprehensive evidence within the category. Unknowns limit a score; they do not prove poor product quality. Community ratings concern support evidence, not popularity or adoption.