Build a source-grounded task brief

Start with the job, not a clever opening

Before writing a prompt, describe the work you need back. “Summarise this” can mean a paragraph for a colleague, an action register for a project or a digest for a new team member. Those outputs need different choices. Name the person using the result and the decision or action it should support. A small, bounded task is easier to inspect than a request to solve everything at once.

Gather the material the answer can rely on

Put the relevant notes, facts and document extracts together. Include dates, units, definitions and exceptions when they affect the meaning. Remove private details you do not need to send to another service. If a file is only mentioned by name, do not assume a model can read it. Either provide the relevant extract or state that the file is unavailable. Keep original sources somewhere you can revisit.

Separate facts from instructions

Use a distinct source section in the brief. Tell the model to treat it as reference material rather than a new set of instructions. This is useful organisation, but wording alone does not guarantee that every model will behave as intended. Read the result and check for material that appears to have come from instructions inside the source instead of your actual task.

Describe the shape of the answer

Choose a format that fits the work. An action register might need task, owner, deadline and evidence columns. A document digest might need an overview, limitations and unanswered questions. Say how unknown values should appear. “Not specified” is often more useful than a confident guess. If exact machine-readable output matters, validate it with a suitable parser before putting it into another system.

Use a small example

Suppose your notes say Mina will send a draft on Tuesday, Leo suggested a call, and the budget still needs approval. A useful brief asks for agreed actions separately from suggestions and open decisions. The expected result contains Mina’s action, marks the call as a suggestion and leaves the budget unresolved. It does not assign Leo a deadline or claim the budget was approved.

Make the checks observable

Decide what you will check before reading the response. In the example, names and dates must match the notes, suggestions must not become commitments, and missing owners must stay missing. Avoid relying on a single impression that the answer sounds polished. You can evaluate factual fidelity, coverage, structure and usefulness separately. Keep the source beside the answer while reviewing it.

Save the process, then adapt it

Use Task Brief Builder to choose a task card and enter your material. Review the assembled prompt, copy it to your chosen AI service and inspect the response. Save a version when a brief becomes useful, but revisit the facts each time you reuse it. A reusable structure is valuable; yesterday’s assumptions or confidential material should not travel into a new task unnoticed.