Skip to content

Quickstart

1. Start

/gru add rate limiting to the public API

On the first run in a project Gru runs the interview: models (stay on Opus/Sonnet/Haiku or change), the minion harness ("default is Claude Code, I suggest tether"), limits for local or rate-limited models, the project's checks and what "done" means, and how you like to work. It then offers a selftest.

2. Get grilled

Gru reads your repo docs first and never asks what they answer. Then at most 7 questions per round, each with a recommendation. It may conclude "don't build this".

3. Review the cards

You see each card's objective, definition of done, size, the model per role, the schedule (sequential where a local or rate-limited model forces it), the estimated cost and the budget. Change a size, a model, a definition of done or the order; your word wins. Then accept.

4. Watch it work

/gru status

A table of cards with stage, models, replans and observations. Guide a card without stopping the others:

/gru steer C2 use the existing Redis client
/gru steer C2 bump

5. Take the result

Gru hands back a structured result: per card status, branch or PR, evidence for every definition-of-done line, and open should-fix items. Verify it or deliver it.

/gru resume    # after a closed terminal
/gru done      # wrap up, clean leftovers, archive