Text briefing | Sources checked October 23, 2026
How to Write Better AI Prompts: Task, Context, and Check
A polished AI draft can still be wrong. Before asking for one, write down three things: the job you want done, the facts it may use, and how you will check the result. Then keep those notes beside the answer.
- For
- Anyone using AI to turn rough notes into a short, shareable draft.
- Use it to
- Trace a draft back to its source before sending, printing, or posting it.
One brief, three parts
Earlier ShortBriefing lessons explored why vague prompts produce vague answers, how to define a useful result, and how to check a draft against its request. Here is a way to put those ideas together for one small task.
Task: Name the artifact, its reader, and its format. Our example asks for a short instruction card for a first-time maker, starting with an inventory line and followed by exactly three numbered steps. It must stay under 85 words.
Context: Supply the actual maker notes and name what is missing. In this entirely fictional Star Bookmark Kit, the parts are one paper strip with a pre-punched hole, two star stickers, and one tassel with a loop. The maker's ordered steps are to put the stickers on the front, pass the tassel loop through the hole, and pull the tassel through its own loop. The notes say nothing about tools, colors, drying time, a completion time, or additional steps.
Check: Decide before drafting what to compare. Does the card list the three parts and quantities? Are the actions in the maker's order? Has it added anything not in the notes? Is it under the word limit? For a real kit, can a first-time maker physically follow it? The last check cannot be settled by text alone.
The actual prompt and response
We sent this prompt once to a Codex model on October 23, 2026. The kit is invented; this is one observed response, not a test of whether this prompt is better than another one.
Prompt sent
Write a short instruction card for a fictional Star Bookmark Kit for a first-time maker. Use only these maker notes: The kit contains one paper strip with a pre-punched hole, two star stickers, and one tassel with a loop. The maker's steps are: (1) place the stickers on the front of the strip; (2) pass the tassel loop through the pre-punched hole; (3) pull the tassel through its own loop to fasten it. Start with 'In the kit:' and list the supplied parts, then give exactly three numbered steps in the stated order. Keep the whole card under 85 words. Do not add tools, colors, drying time, hole punching, or extra steps. If something necessary is unclear, add a separate 'Check with maker:' line after the steps rather than guessing.
The model returned this card, reproduced without edits:
Observed response
In the kit: one paper strip with a pre-punched hole, two star stickers, and one tassel with a loop. 1. Place the stickers on the front of the strip. 2. Pass the tassel loop through the pre-punched hole. 3. Pull the tassel through its own loop to fasten it.
Check the answer against the notes
The inventory line matches all three supplied parts and quantities. Step one, step two, and step three follow the maker's notes in order. The card adds no tool, color, drying claim, hole-punching instruction, completion time, or extra step. It has 49 whitespace-separated words, including the step numbers, so it is under the requested 85-word limit.
The response contains no “Check with maker:” line. That does not prove the directions are ready to print. No physical bookmark was assembled for this example. A real maker would still check whether the loop fits the hole, the stickers fit on the front, and a first-time maker understands the fastening step. If that test exposes a gap, revise the source notes and the card, then check again.
OpenAI's current prompting guidance recommends clear, specific requests with enough context, followed by reviewing and refining the response. Its accuracy guidance warns that confident answers can still be wrong. Anthropic's guidance on reducing hallucinations likewise recommends grounding claims in supplied material and checking them. None of those sources validates this fictional kit or guarantees an accurate response.
One step you can use
Write three lines before prompting: Task, Context, Check. Keep your notes beside the draft and point to a source for every fact before sharing.
For a small, non-sensitive task, specify what to make, for whom, and in what format; include the source facts and unknowns; then check each claim against the source and ask the real owner to test what words cannot.
This lesson in three formats
Read or watch the same example
Read the companion Substack letter or watch the narrated lesson on YouTube. Both use the same fictional Star Bookmark Kit and the same checked drafting method.
Limits
This is a fictional teaching example, not a physical product test or a measured model comparison. One observed Codex response is shown verbatim; another run may differ.
Disclosures
AI assisted research and drafting of this lesson. The kit notes, teaching method and final wording were edited and checked for this release. The companion video uses synthetic narration and original project graphics.
Corrections
No corrections have been recorded. See the corrections process.
Original sources and limits
Check the guidance
- OpenAI: Prompt engineering best practices for ChatGPTChecked October 23, 2026; page displayed “Updated: 4 days ago.”
Supports clarity, specificity, context and review. It does not supply the fictional kit.
- OpenAI: Does ChatGPT tell the truth?Checked October 23, 2026; page displayed “Updated: 13 hours ago.”
Supports independent verification of important claims. It does not assess this one response.
- Anthropic: Reduce hallucinationsChecked October 23, 2026
Offers provider guidance on grounding and claim checking. It does not validate this fictional example or guarantee an outcome.