Better prompts come from better briefs
There is no magic phrase hiding in the last line of a prompt. Describe the job, supply the needed context, set the limits and say what a usable answer looks like. Then revise the part that failed.
A complete workshop source card, acceptable draft, repair table and second-source test make revisions checkable.
Diagnose an invented detail, missing fact or wrong deliverable before changing the prompt.
Keep in mind: Clear instructions can reduce ambiguity. They cannot supply missing evidence, guarantee truth or erase a model’s limits.
In this article6 sections
Editorial note: These prompts are test cases, not formulas that guarantee a correct answer. Results vary by model and context, and any claim that matters still needs to be checked outside the chat.
A confident answer can expose a vague brief
Suppose you ask an assistant to announce a workshop and receive a polished notice with a room number, registration link and promise of refreshments. If none of those facts came from you, the notice needs repair before anyone sees it. Calling the model an expert or asking it to try harder does not supply the missing information.
This exercise starts with a small source card and a definition of an acceptable result. It then changes one instruction at a time. The notices and failure examples are editorial teaching material, not recorded outputs from a named model. You can run the exercise yourself and compare the answer with the supplied facts.
The source card is shorter than the prompt
Our fictional event is a free introduction to spreadsheet formulas on October 12 at 2 p.m. It lasts one hour. Participants should bring a laptop. The room and registration process are not confirmed. The source gives no year or time zone, so a public announcement would need those clarified where the audience could be confused.
For this practice task, ask for a short draft for a local noticeboard. Do not demand a word count that forces the assistant to manufacture benefits or details. A concise draft can be useful while still marking the details the organizer must resolve.
Draft a short notice using only this source card. Audience: beginners at a local workshop. Include topic, price, supplied date and time, duration and what to bring. Put unconfirmed details in a separate “Organizer to confirm” list. Do not invent a room, year, time zone, booking link or refreshments.An acceptable draft and its unfinished work
One acceptable draft reads: “Learn the basics of spreadsheet formulas at a free, one-hour workshop on October 12 at 2 p.m. Beginners are welcome. Please bring a laptop.” The organizer’s list should still include the room, registration arrangements and any missing date or time context needed for publication.
This is not ready to distribute merely because every sentence is supported. Readers still need to know where to go. That distinction matters: a prompt can produce an accurate incomplete draft, while the publishing decision remains blocked by missing information. The right next action is to ask the organizer, not ask the model to guess.
Use a repair that matches the failure
Compare the response with the source card before rewriting the entire prompt. If the problem is one unsupported detail, preserve the rest of the request and change the instruction about missing facts. If the facts are correct but the notice is too technical, adjust the audience instruction instead.
Scroll the table sideways to see every column.
| Observed problem | Likely gap in the brief | Targeted revision |
|---|---|---|
| Invented booking link | No rule for missing information | Require unconfirmed details in a separate list. |
| Long explanation of spreadsheet software | Deliverable not bounded | Ask for a short event notice, not a tutorial. |
| Laptop requirement omitted | Required fields not explicit | List the facts the notice must preserve. |
| Room marked confirmed after a guess | Evidence and suggestion mixed | Allow suggested wording, but forbid new event facts. |
Retest the change on a second source card
Now change the source to a paid workshop and explicitly remove the laptop requirement. A useful prompt should preserve the new price and stop repeating the old requirement. This is a small check against a prompt that works only because the first example was convenient.
Keep the original and revised prompt, both source cards and the outputs. Record which error the revision was meant to fix. A single success does not establish reliability; it does make the comparison less dependent on whether the newest answer feels nicer to read.
Know which problems wording cannot fix
A request cannot make an unavailable document available or turn uncertain information into a verified fact. When the source is missing, retrieve it or leave the claim unresolved. When the task needs calculation, check the calculation independently. When an action would affect someone else, obtain the required authority outside the prompt.
Treasury Board’s guidance and NIST’s risk profile supply background on output checking and confabulation. They do not endorse this particular brief. Our contribution is the repair procedure: identify the failed requirement, change the relevant instruction, and test whether the change solved that problem without introducing another.
AI answer evaluation worksheetSeparate a wording improvement from a result that actually meets the requirements.
Continue with the original sources
These claim-relevant primary and first-party references support the reporting above. Open them for technical detail, current requirements and subsequent updates.
- canada.caTreasury Board: Guide on the use of generative AI ↗Federal workplace guidance on checking outputs and managing information. Its institutional requirements are not a universal rule for every Canadian business.
- developers.openai.comOpenAI: Working with evals ↗First-party documentation on defining test criteria, datasets and evaluation runs; not independent evidence that a particular model performs well.
- nist.govNIST: Generative AI Risk Management Profile ↗Risk-management background, including confabulation and information integrity. It does not certify the examples or prescribe our scoring thresholds.
Finished reading? Save that here without waiting for a timer.
Rewritten throughout on September 21, 2026, with a complete worked method, explicit evidence limits and checked primary sources.
See something we should fix or clarify? Read the corrections policy or tell the newsroom. Material changes are noted here.
