Superlite is a Claude Code workflow 3.9x faster than Superpowers
Superlite is a Claude Code workflow that forks the Superpowers plugin, cuts the review checkpoints to one, runs up to 18 agents in parallel, and reports an 8.8-minute run on a benchmark that took Superpowers 34.4 minutes.
Image: GitHub
Why it mattersA team already using Superpowers for autonomous Claude Code runs can swap in a workflow whose author measures a 3.9x speed-up at identical test coverage, with the full project and scoring script published for anyone to repeat.
A Claude Code user tired of waiting on multi-step runs has published Superlite, a fork of the Superpowers workflow that trades Superpowers' per-step human review for a single approval gate and runs up to 18 agents in parallel. The repository was created two days ago and has gained 104 stars under an MIT licence.
The author, who goes by iha2 on GitHub, credits Jesse Vincent's obra/superpowers plugin for the plugin layout, session-start hook, and lifecycle skills, and says Superlite tunes those ideas for a different job. Where Superpowers keeps a human close to every step, Superlite is for a user who wants to hand the agent a specification once and come back to finished, tested work.
What the benchmark measures
The bench entry in the repository compares Superpowers 6.4.1 and Superlite 0.1.0 on the same nine-module project, headless on Opus, scored against a hidden 89-test suite. Both runs passed 89 of 89 tests. Wall-clock time came in at 34.4 minutes for Superpowers and 8.8 minutes for Superlite, which the author records as 3.9x faster. A note in the report corrects an earlier version of the numbers, which overstated both costs because of a scoring bug.
The author also publishes a longer run on a sitegen and minisheet pair with 177 hidden tests over three runs per arm, and states that those runs are not directly comparable because they happened at different times. The report ends with the caveat to treat the gaps as indications rather than statistical proof, and notes that Superlite can cost more than Superpowers and scores lower on an extra set of robustness probes, where Superlite averages 4 of 7 against Superpowers' 5.7 of 7.
How the one approval gate works
A user starts the workflow with a specification and receives a plain-language review plan in a .review.md file, which the README describes as two pages with diagrams. Nothing is built, dispatched or written until the user replies approve. After approval the coordinator rewrites the plan into parallel waves, each wave a group of tasks that can run at the same time without stepping on each other, and dispatches them.
The model router picks the cheapest capable model for each task: mechanical work goes to Haiku, more complex work to Sonnet, and the final review and orchestration stay on Opus. Up to 18 agents can be in flight at once, against Superpowers' three or four. An autonomous mode skips the review step for users who want to leave a run going overnight.
Source
iha2/superlite on GitHub. Parent workflow: obra/superpowers.
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.
