Superlite is a Claude Code plugin that turns a specification into finished, tested code with minimal human intervention by planning for parallel execution and enforcing rules in code rather than prompts. It replaces the typical cycle of many small approval gates and strictly sequential tasks with a single approval step (except for destructive or security-sensitive actions), a task graph that lists file ownership and interfaces, and concurrent dispatch to specialized agents: Opus as coordinator, Sonnet for most builds and reviews, Haiku for small mechanical tasks, a fresh Opus for final review, and tiered fix-only agents. Hooks prevent unwanted behavior (for example blocking the coordinator from editing code or turning blocking test runs into background jobs), and every run logs tokens, cost and time so future runs predict performance and respect soft budgets. Users choose whether to review the plan or let the AI orchestrate everything; the system never pushes to main without a human.
Benchmarks show substantial speedups on a nontrivial test project (a static site generator plus a spreadsheet engine scored by 177 hidden tests): a headless run completed in 16.1 minutes with 21 agents involved and up to 18 working concurrently, passing all 177 tests and asking the human zero questions. On the same project, Superpowers ran 33.6 minutes (stock) and 23.5 minutes (orchestrated), so Superlite cut wall-clock time by roughly 52% versus stock and 31% versus an orchestrated workflow while preserving test correctness. Installation requires Claude Code, git, Python 3.10+, and bash; progress and decision logs live in .superlite/ for audit and follow-up.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.