Superlite reports a 16.1-minute autonomous build, and says hooks had to replace prompt rules to get there

Wogu built superlite as a Claude Code plugin for autonomous development, forked from the superpowers workflow (code at iha2/superlite). It asks one batch of questions, gets one approval, then runs a task graph in which each task lists the files it owns and the tests it must pass. Independent tasks run at the same time, up to 18 agents at once in the latest run.

On a site generator plus a spreadsheet engine scored by 177 hidden tests, the latest run took 16.1 minutes, passed all 177, used 21 agents and stopped to ask a human nothing. Stock superpowers averaged 33.6 minutes over three runs (21.2, 31.9 and 47.6), and an orchestrated variant averaged 23.5. Both passed 177 of 177. The 16.1-minute figure is a single run of the current build, and the author calls the comparison "a strong hint rather than peer-reviewed science." The repository README adds two caveats the post leaves out: superlite can cost more than superpowers, and it scored 4 of 7 on extra robustness probes against about 5.7 of 7. We did not run the benchmark harness in the repository.

The design detail worth copying is how it enforces its own rules. Wogu reports that telling the coordinator in bold capitals not to edit code did not work: "The coordinator will read it, nod, and edit the code." A hook now blocks the coordinator from editing project code, another turns raw test commands into background runs, and a plan must pass a lint step before any work is dispatched. Decisions the agent would normally ask about are logged as rulings with the cost of being wrong, so the person returning can review the risky ones first.

Why it matters: The README reports lower robustness scores alongside the speed, and the author does not say what causes the gap, so treat the trade as open. The transferable point is narrower and better supported: if you run an orchestrating agent, put the rules you cannot afford to have broken in hooks, because the author found that a prompt instruction alone did not hold.