How to give AI the context it needs

“Write a project update” leaves important decisions open. Who will read it? Which facts matter? Is the writer allowed to suggest a deadline? A useful prompt answers these questions before asking for prose.

In this exercise you will build a reusable task brief. You need a text-capable AI chat tool and a document for saving the brief. No coding or upload is required. Use the synthetic notes below; hands-on timing remains unmeasured.

1. Identify the missing context

A vague prompt is:

Write an impressive update about our pilot. Make it sound positive.

It supplies no evidence. A confident-sounding response cannot establish what happened. Instead, prepare this brief from original fictional facts dated 6 October 2026:

Context means relevant information, not every document you can find. Extra material creates more items to review. Preserve source labels and dates when they affect meaning.

2. Send a bounded prompt

Task: Draft a pilot status update for the branch pilot manager.
As-of date: 6 October 2026. These are synthetic practice facts.
Evidence:
P1: Three of five branches completed the checklist.
P2: Two await access approval; completion dates are unknown.
P3: Dana asks the access team for an update on 7 October.
Output: Status, Next action, Open question; fewer than 80 words.
Constraints: Use only P1–P3. Do not invent a completion date or claim
success, failure, or schedule adherence. Ask a question where facts are missing.
Review support: After the update, list the note IDs supporting each factual sentence.

Compare the answer with the vague prompt if you wish, using separate conversations so the first answer does not inherit the second prompt’s facts. Save both prompts and outputs; do not interpret a single comparison as a general measure of model quality.

3. Review what specificity changed

An illustrative acceptable update is:

Three of five branches have completed the checklist. Two await access approval, with completion dates still unknown. Dana will request an access-team update on 7 October. When will the remaining branches receive access?

The first two sentences trace to P1 and P2. The next action traces to P3. The final question seeks information rather than presenting an assumed date.

If the response includes “approval will arrive tomorrow,” ask it to remove the statement because no supplied note supports it. A model’s own source-label list is also something to check: naming P2 does not make an invented date part of P2.

4. Make the brief reusable

Save this template, replacing every bracketed field before sending it:

Task: [specific deliverable]
Audience and intended use: [reader and decision]
Period and scope: [date, team, units if relevant]
Inputs: [labelled facts, documents or data]
Known gaps: [missing information]
Output: [format and length]
Constraints: [what must be preserved or not inferred]
Review criteria: [how you will check correctness]

The business-task worksheet includes these fields. Do not retain the pilot’s dates or counts when adapting it to another task. Review workplace data permissions before pasting information into a tool.

Common mistakes and next exercise

Giving a role such as “expert consultant” does not supply missing facts. Add the evidence instead. If a long prompt contradicts itself, remove conflicting requirements and clarify which source is authoritative.

Try removing P3. An acceptable answer should leave the next action unresolved or propose it explicitly as a suggestion, not retain Dana’s action as confirmed. Continue with How to verify an AI-generated business analysis to review a response sentence by sentence.

Source: original fictional pilot brief; outputs are editorial illustrations. No named service run or measured comparison is claimed.