prompt-engineering

A model that reasons natively does not need to be taught how to think. It needs to be told what done looks like, and left alone. Instruction text that explains, encourages, or supervises costs more than it buys — the model already does those things, and your text competes with its own judgment.

Prompt text lives on four surfaces, and the same words behave differently on each. A rule that helps in a request can tax every turn for months as persistent context. Identify the surface before writing.

On encountering a defective prompt

Do not read past it. Prompt text that violates these rules is found far more often than it is written — in a skill, a tool description, a standing instruction file, a subagent brief.

• Already editing that file — fix it, in the same change.
• Not editing that file — name the defect and the fix, once, in a sentence. Then continue the task you were asked to do.
• Flag what changes behavior, not what offends taste. A stale fact, a contradiction between two rules, a filter that suppresses recall, a rule the model already follows — these cost output. Wording preferences do not.
• One flag, then proceed. A task does not become a prompt audit because a prompt was involved. Raising the same class of defect repeatedly in one session costs more attention than the defect does.