gauntletlisted
Install: claude install-skill yoelgal/agent-tools
# Gauntlet - grill the idea, hand off one loop prompt
A Gauntlet Loop is a way to run one long autonomous build: hand the lead agent a destination
plus a reference artifact worth beating, then let builder-critic loops chase that reference
until the human calls the run (step 3 owns the operative mechanics). The method comes from Matt
Shumer's Claude-of-Duty run (credited in `NOTICE`); it generalizes past games to any software
the loop can inspect: a web app, a CLI, a design system, an API.
This skill's whole job is the prompt. It grills the idea until the prompt can be written, writes
it, and hands it to the user to paste into a **fresh** session - fresh because the run wants its
entire context budget, and because this conversation's history would leak your framing into
decompositions the lead agent should own. When the conversation drifts toward building the thing
here, hand the prompt over instead - a gauntlet run inside a half-spent session is the method minus
its advantage.
Read `.better-dev/overrides.md` first when a repo is wired; a project override wins over any
default below.
## 1. Grill until the slots are settled
Six slots decide the prompt. Batch the questions the conversation has not already answered; propose
a default wherever one is sensible and let the user correct it - the grill is done when every slot
is either filled or deliberately delegated to the run, not when the user stops answering.
| Slot | What settles it |
|---|---|
| Goal | The destination in