How to Give AI Better Information Before It Starts

When an AI answer is vague, the usual advice is to write a better prompt. Sometimes the prompt is not the main problem.

The AI may be missing the policy, notes, examples, figures or decision that a person would need to do the same job. No clever instruction can repair a missing fact.

Before you ask AI to begin, build a small source pack. Give it the right information, explain which source matters most and tell it what to do when something is missing.

Start with the job, not the files

Do not upload everything simply because it is available.

First state the outcome. You might need a summary for senior managers, a reply to a customer or a comparison of three suppliers. The task tells you which information belongs in the source pack.

A useful source earns its place because the AI needs it to understand the job, support a claim or follow a rule.

Separate facts, rules and examples

Different sources do different jobs.

• Facts tell the AI what happened, who was involved or what the figures show.
• Rules set boundaries, required wording or a process that must be followed.
• Examples show the expected style or shape of the result.

Label them clearly. If you attach a current policy and an older example, say that the policy controls the facts while the example shows tone only.

Without that instruction, AI may copy an outdated detail from the document that merely looks most similar to the requested output.

Remove material that creates noise

More context is not always better context.

Duplicate drafts, irrelevant email chains and old versions can pull an answer in different directions. A long file can also hide the one paragraph that changes the decision.

Before uploading:

1. Remove duplicates.
2. Give files clear names.
3. Mark the current version.
4. Remove personal or confidential information the task does not need.
5. Note any source that is incomplete or unconfirmed.

If two sources disagree, do not quietly choose one. Tell the AI to report the conflict.

Explain the order of authority

AI needs to know which material wins when sources conflict.

For example:

Use the policy dated 2 September as the source of truth for rules. Use the meeting notes for agreed actions. Use the previous email only as a style example. If the sources disagree, stop and list the conflict rather than choosing one.

That short instruction can prevent an old document from overruling a current decision.

Add a missing-information gate

Ask the AI to check the pack before it starts writing.

Review the attached material for this task: [describe the outcome]. List any information a person would need but the sources do not provide. Separate missing facts that block the task from useful preferences that can wait. Do not draft the final output until the blocking gaps are resolved.

This changes the first response. Instead of producing a polished guess, the AI tells you what it lacks.

Keep a short source note

For work that will continue, record what each file is for and when it was checked.

Your note could include:

• File name.
• What it contributes.
• Owner or original source.
• Date checked.
• Current, draft or superseded.
• Restrictions on use.

You do not need a complicated system. A short list is enough to stop future you from wondering why a file was included.

A practical example

Suppose you want AI to draft a staff update about a service change.

The weak approach is to upload six email threads and ask for an announcement.

The stronger approach is to provide the confirmed change, affected dates, approved contact details, current policy and one previous update as a tone example. You then tell the AI not to infer missing dates or promises.

The second pack is smaller. It is also more useful because every item has a clear role.

Check the answer against the pack

Once the draft is ready, ask the AI to map each factual statement back to a source. Then check the important claims yourself.

Look for information that appeared in the answer but nowhere in the pack. It may be a sensible assumption. It may also be invented.

The source pack helps AI produce the work. It also gives you a practical way to review it.