Write Prompts

Use this skill when the artifact is prompt text or prompt-like instructions. The job is not to make the prompt longer. The job is to make the model's task clearer and move non-language constraints into better mechanisms.

Principles

• Assume the model is smart. Explain the task, context, constraints, and output contract; do not teach obvious reasoning.
• Delete before adding. Remove stale cases, duplicated rules, contradictions, and defensive prose before introducing new wording.
• Separate layers: prompt intent, examples, schema, code, tests, evals, tools, loop state, and memory are different surfaces.
• Put deterministic enforcement in deterministic places: types, schemas, tests, linters, validators, scripts, or harnesses.
• Prefer compact general principles over piles of narrow case law.
• When the prompt runs in a loop, design the loop's observations, verifier, state, halt policy, and escalation path too.