An AI tuned Superpowers 6, cutting build time 50% and token cost 60%

Superpowers, a plugin that orchestrates autonomous subagent-driven coding through heavy up-front planning, strict red-green TDD, and dual-axis review of every change, has long been criticized for being slow and token-hungry. Version 6 targets exactly that, claiming roughly 50% faster builds and 60% lower token spend with no drop in output quality. The planned 5.2 release (adding harness support for Pi, Antigravity, and Kimi Code, better behavior on Codex, OpenCode, and Cursor, and model-agnostic skill rewrites) got sidelined when the author gained brief access to Anthropic’s short-lived Fable model and used it to optimize the build loop itself.

The biggest savings came from a few structural changes. Pre-generating a review ‘packet’ — a formatted diff plus metadata via shell script — stopped reviewer subagents from burning tokens running ad hoc git commands, cutting time and cost about 10%. Merging the separate code-quality and spec-compliance reviewers into one saved another 15%. A terse reviewer contract dropped reviewer output 41% with verdicts intact, and a standardized narration recipe cut 54%. Notably, Fable ran an overnight ‘autoresearch’ loop of 25 pre-registered experiments (about $165 in spend, Opus as coordinator) and independently reached the reviewer-merge conclusion the author had jotted down the same night. The loop also disproved some intuitions: capping controller ‘thinking’ backfired, raising turn counts and doubling output.

The write-up is refreshingly candid about measurement discipline. Three of the author’s own benchmarking bugs were caught mid-loop, forcing a headline reviewer-savings number to be corrected from an inflated 74% down to an honest 41%. An early Codex run showed zero improvement — until they found the eval harness wasn’t isolated from the host OS and had been silently benchmarking the old 5.1.0 the whole time; once fixed, the gains held. The episode is a useful case study in both AI-assisted self-optimization and why a real evals suite (open-sourced here) is a prerequisite for trusting any of these claims. Superpowers 6 is available now on GitHub.