Write an AI brief that keeps facts separate from guesses

Choose one output and its acceptance rule

A useful prompt begins with a deliverable someone can inspect. “Help with marketing” has no clear finish. “Write an 80-word listing using these five product facts, then list unanswered buyer questions” does. State the reader, the decision they need to make, the output format and a length ceiling. Treat tone as the last constraint: friendly language cannot repair a wrong size or an invented delivery promise.

Use the exercise below with the fictional Foldwell desk-label pack. Its facts are deliberately small enough to check by hand. The goal is a listing draft, not market research. The process is our suggested editorial method; NIST's voluntary AI risk framework provides broader context for evaluating AI use, not certification of this prompt.

NIST: AI Risk Management Framework ↗

Make a source packet before asking for prose

Number the facts and preserve exact quantities, units and scope. Separate a statement from the evidence that supports it. A note saying “waterproof?” belongs in questions, not facts. Put facts from different dates in separate entries; do not let the model silently reconcile conflicting revisions. Remove private customer details when they are irrelevant to the requested draft.

For this exercise, F1 says the pack contains 12 blank paper labels. F2 says each is 40 × 15 mm. F3 says a black pen is included. F4 says the labels are for indoor use. No evidence is supplied about adhesive, shipping, returns or water resistance. That absence matters: an attractive listing must leave these claims out.

F1: 12 blank paper labels per pack.
F2: Each label is 40 × 15 mm.
F3: One black pen is included.
F4: Intended for indoor use.
Unknown: adhesive, water resistance, price, shipping and returns.

Copy a brief with an explicit boundary

Paste the prompt into a writing assistant you already have permission to use. In KitForma, choose the local AI assistant for a short exercise. The current source limits input to 6,000 characters and describes a small experimental model without live web access. Keep each source packet short and review the response; this guide does not claim a benchmark or a successful run on your device.

The boundary is an instruction for the draft, not a security barrier. Source documents can contain misleading instructions or false statements. Treat quoted material as content to evaluate, and never let a paragraph inside it authorize sending files, installing tools or changing an account. The local exercise requires none of those actions.

Task: Draft a plain English product listing, at most 80 words.
Reader: A person labelling desk cables.
Use only F1–F4 below. Keep numbers and units exact.
Do not add benefits, materials, guarantees or delivery terms that are absent.
Return: (1) listing; (2) claim → fact ID list; (3) missing buyer information.
If a requested detail is missing, say “not supplied”.
Facts:
F1: 12 blank paper labels. F2: 40 × 15 mm each.
F3: One black pen included. F4: Indoor use.

Compare the draft with an expected answer

A defensible illustrative example is: “Label your desk cables with a pack of 12 blank paper labels. Each label measures 40 × 15 mm, and one black pen is included. The labels are intended for indoor use.” Its statements map to F1–F4. This example was prepared for this AI-assisted guide; it is not a recorded model test response.

Reject “12 reusable waterproof labels with fast delivery.” The count has support, but reuse, water resistance and delivery do not. A citation-looking fact ID beside that sentence would not fix it. Break compound sentences into claims: each adjective and promise needs its own support. If an assistant changes millimetres to centimetres, repair the source-to-output comparison before polishing style.

Revise one failure at a time

Keep the first output and annotate the specific defect. “Make this better” obscures what changed. Instead ask for a shorter opening while freezing quantities and removing unsupported promises. Request one revision, compare it with the original facts, and record whether the defect recurs. A different answer on a second run does not establish that the underlying issue is solved.

If a small model ignores the requested three-part format, split the task: first ask for a two-sentence listing, then audit that listing manually. Use the dedicated product-description tool for a concise product draft; use the general assistant for a brief that asks for commentary. Tool-specific output rules may take precedence over elaborate user formatting.

Revise only the listing below. Keep 12, 40 × 15 mm, one black pen and indoor use. Remove every unsupported adjective. Do not add a price, delivery time or guarantee. Return at most two sentences.
Draft: [paste the draft]

Save a small record with the final draft

Store the source facts, prompt version, raw response, corrected draft and unresolved questions together. For a real listing, have the person who owns the product data approve the final factual claims. An assistant's confidence score is not that approval. Repeat the check when the product or source changes; a correct old draft can become inaccurate without any change in the prompt.

Finish this exercise by writing three missing buyer questions and a factual two-sentence description. You pass when all four supplied facts survive intact, no unsupported claim appears, and the unknowns remain visible. These checks are intentionally narrower than legal, commercial or publication approval.