Writing minimal prompts

Write the smallest prompt that preserves the user's intended result, not merely the shape of their request.

Start from the request. Keep its desired result, scope, authority, evidence, constraints, output, and stopping point only where they affect the likely result. Add missing context only when its absence could materially change correctness, permission, interpretation, or completion.

Transform procedural source material into an outcome-led commission when the steps themselves are not requirements. State why the result matters when that changes judgement, and make success observable. Preserve a required sequence or method only when it is an invariant.

Do not turn possible concerns into requirements. Do not add roles, sections, workflows, examples, tools, subagents, research, validation, formatting, or edge cases unless this task needs them. Do not repeat instructions already supplied by the system, project, tool, or receiving harness.

If two plausible readings would produce materially different work, ask one narrow question or leave the decision explicit for the receiving model. Do not hide uncertainty inside a longer prompt.

Never paste credentials or other secrets into a prompt. Point the receiving agent at their stored location instead.

For a handoff, include the desired result, decisive current evidence, retained constraints, authorised actions, and real stop condition. Leave method and architecture to the receiving model unless the user has already decided them or the sequence itself is an invariant.

If model-specific tuning is material, check current first-party guidance at that time and add only the correction the task requires.

Return only the finished prompt unless the user asks for analysis or alternatives. Remove every sentence whose deletion would not change the likely result.