A reusable prompt brief for real workplace tasks

Start with a deliverable, not a personality

"Act as an expert" does not tell an assistant which facts it may use or what a usable answer looks like. For a repeatable task, write a short contract between the requester, the assistant and the person who will approve the result. Keep that brief beside the source material, not inside an employee's memory.

Copy this brief

Task: [specific deliverable and intended audience]
Decision owner: [person who approves the work]
Approved inputs: [files or facts, with dates and source labels]
Boundaries: [private data to exclude; actions not authorized]
Required output: [format, length and required fields]
Uncertainty: flag missing evidence; do not fill gaps with inventions
Review: [checks a human must complete before use]

Use headings or delimiters to distinguish instructions from material to summarize. Include only context that helps the task. A long transcript of unrelated discussions can obscure the current requirement. These are design habits, not a promise that prompting alone makes a model reliable.

Worked example: a delayed order

Use this synthetic fixture: order SAMPLE-42 was due October 20; the warehouse confirmed October 23 as the revised dispatch date. No carrier delivery estimate exists. The customer should receive a courteous update, but no refund or discount is authorized.

The task is to draft a 120-word email. Require the new dispatch date, an apology, a distinction between dispatch and delivery, and an invitation to contact support. Explicitly forbid inventing a delivery date or promising compensation. Ask for a second section listing facts that need verification; do not send either section automatically.

A satisfactory draft says the order is expected to dispatch October 23 and that delivery timing will be confirmed separately. A polished promise that it will arrive October 24 fails, even if its tone is excellent. Reject a draft that adds a discount. The reviewer compares every commitment with the fixture before approving it.

Reuse and improve

Save the brief, one approved output and a failing example together. Test the same brief on a normal case, a missing-date case and a contradictory-date case. A useful revision should improve those cases without making the normal one worse. Record the tool and test date; reassess after changing the prompt, model or source process.

Read next: verify AI-assisted work.

Sources and review

Reviewed October 16, 2026. Template and SAMPLE-42 fixture are original EAIRI examples. Primary references: OpenAI prompting guidance and Anthropic context engineering.