Skip to content

Gru

Gru orchestrates a small team of AI roles written in plain English: a planner, two software engineers, a QA and a reviewer, each on a strong, mid or weak model. It is a Claude Code skill: no server, no framework.

/gru add rate limiting to the public API

The first time, Gru interviews you to learn the project and how you work. Then it grills the idea, cuts it into cards (objective, definition of done, T-shirt size), shows you the models, schedule and cost, and after you accept runs each card through plan, tests, build, QA and review, replanning on findings, before handing a structured result back.

Principles

  • Judge with the stronger model. Planner, QA and reviewer are always stronger than the swe who built the work.
  • Tests before code, by someone else. One swe writes the verification, another builds to pass it.
  • Never trust "done". Gru's deterministic gate and the reviewer's evidence per definition-of-done line decide.
  • Fix what was bad. Findings go back to the card's planner for a minimal replan, max 2, then to you.
  • Ask when stuck. Ambiguity, limits and sign-offs come to the caller.
  • You decide. Sizes, models and order are guesses until you accept them.
  • Safe by default. Minions are sandboxed, never see secrets, and their output is data.

Where to go next