Task cards

A minion sees no conversation, so each card is self-contained. Strong-model tokens are best spent here: weak models fail on vague cards. Each card has an objective and a definition of done.

Field	Why
id, title, size (XS to XL), risk, kind	what it is, and which slots run it
Objective	what changes, for whom, why
Definition of done	a checklist; every line names its check (a test, a command, or what the reviewer inspects)
files_owned, depends_on, resources	scheduling: disjoint files, order, rate-limited or local models
needs	services it needs while testing (database, browser); drives the QA sandbox
Context	at most 10 lines of the small whole picture
Reading list	exact files to read; "do not explore the repo"
Conventions	the few project lines that apply, pasted in
Rules	no commit/push, no touching other files, no .env*
Estimate	tokens by stage, and cost if prices are known

"Works well" or "is clean" is not a definition of done. If Gru cannot make a line checkable it asks you.

Sizing

Gru guesses the size, you decide in the preview (or later with /gru steer <card> resize <size>). An XL probably wants splitting. If est_tokens exceeds half the target model's context window (unknown free models count as 32k), Gru splits the card or trims the reading list.

Preview

Before anything runs you see a table (card, size, risk, dependencies, the model per role, harness, the sandbox per stage, estimated tokens and cost), the schedule (what runs together, and why anything is sequential), the budget, and flags (L/XL sign-offs, cards needing services the sandbox cannot give, calibration notes). Change anything; your decision wins.