Ask Bounded Questions and Verify Sources

A vague request invites a vague answer. "Tell me how to make money in retirement" can produce a generic list that ignores health, pension rules, care duties, and the need for income now. A bounded request gives the tool a task, supplies relevant facts, states what to leave out, and asks for uncertainty to be visible.

Prompting is not a magic phrase collection. It is a way to define the work clearly enough that a person can review it. The best prompt for a high-stakes task may include a statement that the AI should not make a final decision.

A practical prompt structure

Use six parts:

1. Role and task: What kind of help do you want?
2. Inputs: Which approved facts or documents may be used?
3. Audience and format: Who will read it, and how should it be organized?
4. Exclusions: What must the tool not infer, decide, or include?
5. Uncertainty: How should missing or conflicting evidence be marked?
6. Review: What must a human check before use?

Here is an illustrative prompt for a low-risk task:

> Help me organize these approved workshop notes into a one-page outline. Use only the notes below. Do not invent attendance, prices, legal requirements, or accessibility features. List any gap as "unverified." Keep the language plain and printable. I will check every date, contact, and requirement before sending.

The prompt does not make the model trustworthy by itself. It makes the boundaries visible and gives the reviewer a checklist.

Give source material a clear status

Separate the prompt into source categories:

• Verified: A current official document, signed agreement, or inspected record.
• Provided but unchecked: Notes, memory, or a draft that may contain errors.
• Questions: Things the tool may help organize but cannot answer.
• Forbidden: Private, confidential, or irrelevant information that must not be entered.

This distinction prevents accidental authority. If you paste a flyer into a prompt, the tool may treat every claim as true. A person still must confirm that the flyer is current and that the organization authorized the content.

Ask for uncertainty

A useful prompt can require a short evidence note. For each important factual claim, ask the tool to mark:

• the source you supplied
• whether the claim is confirmed or unverified
• a question that needs a human check

Do not request fake confidence or a dramatic recommendation. For a health, legal, financial, benefit, or safety question, request a list of questions for the responsible agency or qualified professional rather than an answer disguised as expertise.

Verify at the source

A model may summarize a source inaccurately. Open the source when the claim matters. Check:

• who produced it
• the date and version
• jurisdiction and population
• definitions and exclusions
• whether a later update replaces it
• whether the organization permits the use

Use the official page or office instead of a link supplied only by the chatbot. Search independently or telephone with a number you found through an official directory. A source can be real and still be obsolete.

Use a claim table

For a document with several factual statements, create a three-part review:

• Claim: What exactly does the draft say?
• Evidence: Which original source supports it?
• Decision: Keep, correct, qualify, or remove.

Example: "The workshop opens at 9:30" is a claim. The hall's signed confirmation is evidence. If the confirmation says 10:00, the reviewer corrects the draft. If no confirmation exists, the claim is removed or marked for follow-up.

This process feels slower than generating text. It is often faster than fixing a polished document after a wrong promise has been sent.

Separate a question from a command

AI often follows an implied goal. "Should I volunteer at this organization?" may produce a recommendation based on thin details. A better prompt could ask:

> Compare this organization's written role description with my access needs and time limits. Do not decide for me. Produce questions about supervision, training, expense reimbursement, safeguarding, insurance, and exit. Flag any statement that needs confirmation by the organization.

The result supports a decision without impersonating the organization or taking responsibility for it.

Account for current information

Language models can be outdated. Even a tool with web access may not retrieve the right page or may summarize it incorrectly. For current schedules, prices, laws, benefits, medications, and emergency contacts, use the authoritative source directly.

If the tool offers links, open them and check that they lead to the claimed material. Avoid entering account credentials or private records into a browsing assistant when a simpler official route is available.

Preserve privacy

Write the prompt around the minimum necessary facts. Remove:

• names not needed for the task
• account numbers and passwords
• health details unrelated to the question
• private client information
• identity documents
• addresses or routines that create risk

Ask whether the tool's terms allow the data. If using an organization's account, follow its policy. When in doubt, do not enter the private material. Use a redacted example instead.

Exercise: Rewrite a vague prompt

Start with: "Help me plan something useful after retirement."

Rewrite it using the six-part structure. State one audience, one permitted set of notes, one task, one privacy limit, one uncertainty rule, and one human review step. Do not aim for a perfect prompt. Aim for a prompt whose output can be checked.

Then ask yourself:

• Could the tool invent a fact that I would accidentally send?
• Does the prompt ask for a decision that belongs to a professional or agency?
• Have I included more personal information than necessary?
• Is there a manual path if the AI result is poor?

A fictional review

Ravi, a fictional retired bookkeeper, wants to draft a short guide for a community group. He supplies the group's approved contact details and a verified schedule, excludes private member information, and asks the AI to note any missing policy. The draft looks polished.

During review, Ravi notices that the tool added a fee that does not exist and omitted a required accessibility contact. He removes the fee, confirms the contact through the organization, and changes the wording so no one is promised a service outcome. The final document is not accepted because AI wrote it smoothly. It is accepted because the sources and reviewer made it accurate.

What to carry forward

A bounded prompt defines permitted input, exclusions, uncertainty, and review. Verification happens at the source, not inside another model answer. The next chapter applies these habits to accessibility, because a tool that cannot be operated safely or comfortably is not useful merely because its output is impressive.