masterprompt

Run reconnaissance at the source first, then draft from the masterprompt template with a one-sentence mission, falsifiable done-conditions, dated ground truth, explicit scope-out, and landmines. Finish with a lint pass and a red-team review before delivering.

A masterprompt is a self-contained brief handed to a fresh executor who then completes the work without asking what you meant. Use it for multi-step work crossing a context boundary; skip it for tasks that fit in the current session or are still being scoped.

Each done-condition must name the observation that would prove it false, made on something the author does not control. Replace vague targets like "improve X" with the exact command or query whose output settles the question.

Write instructions in English, since multilingual models reason best in their dominant training language, but keep data literals like Hebrew UI strings in their original script inside backticks. Always state the output language and register explicitly.

Skip it when the work fits in the current session, when the ask is a question, when scope is still forming, or when you need a task-level plan for already-scoped work. A masterprompt wraps work; a plan nests inside it.

No. Secrets are named and located, never valued, and personal data is referenced by query rather than pasted. Where a human must supply a secret, they set it directly and the executor verifies through an observable side effect.