A single perfect prompt can't win — the failing repros only appear after you build and grade. The skill is the loop: read the machine's real failures, name the precise rule that flips each one, and verify before you card. A naive operator re-pastes 'handle all edge cases' and never moves; a skilled one closes it in two.
Drive it yourself
real sandbox · real Kimi rebuilds
We open a Daytona sandbox and grade a failing seed build. You read the repros, write a patch spec, declare which checks it'll flip, and run — Kimi rebuilds in that same sandbox and re-grades. Re-run on green to verify, then card.
Targeting — flip exactly what you declareConvergence — green in few turnsVerification — re-run before carding
step 1
You're handed a failing build
It's already running but four hidden checks are red — each with a live repro you couldn't have known up front.
step 2
Read a repro, steer the build
Each turn you declare which checks you'll flip and name the precise rule. Kimi rebuilds in the same sandbox; the log updates.
step 3
Verify, then card
Re-run on green before you declare done. The score rewards fewer, sharper turns — and penalizes carding on faith.
Or watch how it's done
two real captured loops · press play
Same seed, same four red checks: a skilled operator flips exactly what each turn declares and walks the ball to the cup; a naive one re-pastes generic prose and never leaves the tee.
Skilled operator
Reads the repros, steers precisely
—
check
seed
patc
patc
veri
card
⛳
“Build an LRU cache with get/put; evict least-recently-used when full.”
Naive operator
Re-pastes 'handle all edge cases'
—
check
seed
patc
patc
card
⛳
“Build an LRU cache with get/put; evict least-recently-used when full.”