AI Agent News/stories/superlite-a-claude-code-plugin-for-unattended-builds-from-a-spec.md

Frameworks

Superlite, a Claude Code plugin for unattended builds from a spec

Superlite is a Claude Code plugin that plans a build, runs independent tasks in parallel on the cheapest suitable agent, and reviews the result with only one upfront approval. Its author reports finishing a 177-test benchmark project in 16.1 minutes, compared with 33.6 minutes for stock superpowers.

Source: contextfirst.blog/superlite-hand-an-agent-a-spec

Iheatu Chris Wogu has released superlite, a Claude Code plugin that plans a build, splits it into tasks that can run side by side, assigns each task to the cheapest agent able to handle it well, checks the seams between tasks and reviews the result. It is meant for autonomous development, where a user hands over a spec and returns later, unlike superpowers, which keeps a human close to each step. Superlite borrows its plugin layout, session-start hook and lifecycle skills from superpowers, but the author advises against enabling both in one session because they share skill names and inject context at startup.

For a new project, superlite asks one batch of questions and asks for one approval of the design and plan, then only stops for destructive, outward-facing or security-sensitive actions. Opus coordinates without editing code, Sonnet builds most tasks, Haiku builds small mechanical tasks, and a fresh Opus reviewer returns findings without fixing them. Hooks enforce these rules: one blocks the coordinator from editing project code, one moves raw test runs into the background, and one refuses to kill processes by name.

On a benchmark project of a static site generator and spreadsheet engine, scored by 177 hidden tests, the author reports a 16.1 minute wall-clock time with all tests passing. That compares with 33.6 minutes for stock superpowers, a mean of three runs, and 23.5 minutes for orchestrated superpowers. The author describes these as single runs on one machine, a strong hint rather than rigorous evidence. The plugin installs from the iha2/superlite repository on GitHub.

Why it matters if you build agents

If you build agents that run long unattended coding jobs, the post shows a pattern worth copying: split work along interface boundaries, run independent tasks at once, and enforce important rules with hooks instead of prompts. The speed figures come from the author's own benchmark runs, so test the plugin on your own project before relying on its timings. The author also advises against running superlite and superpowers in the same session.

Get stories like this every weekday

One short email with the day's agent tooling news.

Subscribe