Turn a Vague AI Prompt Into a Reusable Brief

Before you start

You need a plain-text editor and the fictional inventory below. You do not need an AI account, a paid plan, an API key or access to personal files to complete this exercise. It is a writing-and-checking exercise, not an app walkthrough. No assistant or model version was tested.

Anthropic’s prompting guidance recommends specific instructions, relevant context, smaller steps for complex requests and follow-up feedback. The reusable brief below is our own example of applying that advice. It does not establish that a longer prompt will produce a more accurate answer.

1. Define a task small enough to check

Start with the vague request: “What should we bring to the garden?” It leaves the occasion, available equipment and desired answer open. We have not submitted that request to a model and are not presenting a simulated response as an actual result.

For this exercise, the task is narrower: turn an existing inventory into a packing checklist for a fictional volunteer. The inventory contains one watering can, two hand trowels and three pairs of gloves. The checklist should include every listed item and preserve those quantities. It is not advice about what a real garden needs.

2. Write a reusable brief

Copy the following fields into a note. Keep the field names when reusing it, but replace the task and source facts together. Separate the material the answer should use from the instructions about its shape.

• Task: Turn the supplied inventory into a packing checklist for a fictional garden volunteer.
• Source facts: watering can — 1; hand trowel — 2; pair of gloves — 3.
• Boundary: Use only those items and quantities. Do not add equipment, advice or assumptions. If information needed for the task is missing, identify the gap instead of supplying a fact.
• Output: Exactly three checklist lines, one per inventory item, in the same order. Each line must contain the item and quantity. Add no introduction.
• Check: Before returning the checklist, compare every item and quantity with the supplied inventory.

3. Write a reference answer before judging another answer

A hand-written reference makes the success conditions visible. For this example, it is: “1 watering can”; “2 hand trowels”; “3 pairs of gloves”. These are reference lines written for this exercise, not an assistant’s output.

Check four things: three lines, no duplicate item, no item outside the inventory and no changed quantity. A polished sentence is not a substitute for those checks. A self-check requested inside a prompt still needs an outside check by the reader.

4. Reuse the structure and record discrepancies

Save the brief with its inventory and reference answer. To adapt it, change the source facts, the task and the expected answer together. For a reading summary, for example, the source field would contain the text to summarize and the output field would define the requested structure. That is a different task and needs its own checks.

If you later choose to try the garden exercise in an assistant you already use, keep it free of private information. Record the app, visible model label if available, date, exact prompt and exact response. Compare the response with your reference. We have not carried out that optional model test; availability and account limits are outside this exercise.

If the result does not match

An extra item fails the source boundary even if it sounds useful. A changed quantity fails the inventory check even if the format is correct. A missing item fails completeness. Identify the specific mismatch, edit one instruction and keep both versions if you test again. Do not silently replace a failed result with a successful-looking example.

If the task itself needs facts not in your material, gather and verify those facts first or narrow the task. Repeating an instruction cannot supply missing evidence. For other subjects, a short checklist may miss important errors; design the checks for the actual question.

What was checked, and what remains unknown

On 1 October 2026, an automated local Python check compared three hand-authored fixtures with this inventory. It accepted the reference, flagged a substituted fertilizer item and flagged three trowels where the inventory says two. No network request or model invocation was used. The test checks the example’s rules, not AI behavior.

The expected result is a saved brief plus a reference you can inspect. We have not measured improvement over a vague prompt, time saved or any model’s accuracy. An assistant comparison remains unperformed.