One install wakes N sub-agents on your task — 100 by default, 16 with --quick. Each gets the identical brief and a different strategy card, then they attack, defend and judge each other in a single-elimination bracket until one solution survives.
install
Detects Claude Code, Cursor, Codex, Qoder and more, then writes the skill into the right folder. Needs Python for the bracket bookkeeping. Add --dry-run to see the plan first.
npx skills add arm-pz/do-better
how a run works
Type /do-better <task> — or just say the last answer was bad, and it asks once before spending anything. Every agent gets the identical brief; only the strategy card differs.
Size the tournament. A fixed seed replays the same bracket.
Bookkeeping runs through bracket.py, never by memory.
Every round: both attack, both defend and revise, a judge scores, the loser is eliminated. With a rejected answer on file, one final blind check closes the run.
house rules
It runs the bracket and nothing else: no solutions, no attacks, no judging, no picking winners. If a judge fails twice, a fresh one runs the match — the verdicts stay out of its hands.
Every competitor reads the identical brief built from one task file. No hints to a favourite, no paraphrasing — the only difference between 100 agents is the card they were dealt.
Judges read both revised solutions in full, verify each "fixed" themselves, and score against a written rubric. Length is not quality, and the record says so.
Free, open, and it runs entirely inside the agent you already use.
Install do-better Read the source