Write prompts that make the task unmistakable

Replace adjectives with requirements

“Write a great answer” leaves the standard undefined. A more useful instruction identifies the audience, the question to answer, the evidence to use, and the expected output. For example, ask for a short integration note that names required parameters, explains an error case, and links each factual claim to the supplied reference.

Anthropic’s prompt engineering overview recommends establishing success criteria and an empirical way to test them before tuning a prompt. Apply that principle by drafting a few realistic inputs and describing what a useful response must include.

Give each part one job

• Goal: State the task in plain language.
• Context: Supply the relevant document, data, or conversation.
• Constraints: Specify boundaries, such as using only supplied evidence.
• Output: Describe the length, organization, or required fields.
• Fallback: Explain how to handle missing or contradictory information.

Examples are especially useful when a style label could mean several things. Show a concise response with the structure you want, and choose examples that reflect the real task. Keep placeholders clearly distinct from instructions so a sample value does not become a recurring factual claim.

Revise the cause of the error

If the model omits a required field, clarify the output requirement or move that requirement into a supported schema. If it invents details, examine the source context and uncertainty instructions. If it produces an answer for the wrong audience, make the reader and purpose more explicit.

Change one meaningful element at a time and compare the result across the same cases. Preserve the previous prompt so you can reverse a change that improves one example while damaging several others. Add difficult new cases when real usage exposes a gap.

Keep reusable prompts close to their evaluation notes. Record their intended task, expected inputs, version, and owner. That makes a prompt library easier to maintain than a folder of attractive but untested snippets.

For machine-readable responses, continue to LLM prompt contracts and schemas. For user-facing exchanges, apply the instructions within a deliberate conversation design.