lightarktech Never launch "until it's done". "Done" is not observable; loops without observable exits run forever, burn budget, and drift. Fill this contract first — every field, before launch:
GOAL One sentence: take X from state A to state B.
DONE Observable end state ("tests green + lint clean + PR open").
Banned: "done", "polished", "production ready", "good enough".
SCALE ≤100 turns, ≤5 tasks. Hitting a cap = save progress, report, split.
STOP Reached DONE / hit a cap / stuck needing a human → stop and report.
RETRY Not DONE and moves remain → read the last failure, change ONE
variable, retry. Never shotgun multiple changes per iteration.
VERIFY How it checks its own work: tests, an independent evaluator,
or adversarial review (max 2 rounds).
FIRST Context + what to check before starting — a fresh session
knows nothing about this conversation.
RED LINES What it must never touch (production, user data, permissions,
anything irreversible).
Trust mechanisms, not luck: the point is to put unattended work inside a box that is observable, correctable, and reversible.
Success criteria
A stranger could run the loop from the contract alone. The loop cannot run forever — every path leads to DONE, a cap, or a report.