Start here

You do not need a terminal, an account or any setup. Work in an ordinary AI chat that accepts text.

What to paste: a description of the form (name and purpose), the field labels or a short transcript, which fields are required or optional, the agent tool you plan to test described in your own words, whether you can draft-save, preview or screenshot manually, the most sensitive fields, and any deadline or cost of sending wrong information.

Where to paste it: in a normal chat, together with the full prompt appended below.

How to check the output: the reply should be a Markdown brief with scope and boundary, a sample-data table, numbered checkpoints, a verification table with the agent's column left blank, stop rules, review checks and open questions. Every sample value must be labelled SAMPLE and every unknown must sit under open questions, not be filled in.

What this prompt does

It converts a form you describe into a written test plan for an agent or browser helper that fills fields. The plan covers fake sample data, what evidence to capture manually at each checkpoint, a field-by-field comparison table you complete during the test, and stop rules for submission, passwords, codes and CAPTCHA bypass.

Sensitive categories are marked HIGH-CAUTION. The prompt asks the assistant to describe what to observe, not to claim the agent can or cannot perform an action.

Preparing your input

If you leave out required fields or the sensitive-field list, expect them under open questions. Placeholders you should not paste: real addresses, identity numbers, medical details or account numbers. If you already pasted them, ask for a placeholder substitution.

Permissions and limitations

• The brief is a checklist, not an action. It never authorises submission, payment, account creation, term acceptance, message sending or document upload.
• You perform any live step yourself after review.
• It gives no legal, medical or financial advice about the form content and never suggests disabling security features.
• It does not prove what the tested agent can do; it only structures your observation.
• Documentation only: source reviewed; runtime not tested.

FAQ

Can the prompt submit the form for me? No. It only produces a written brief; every live step stays with you.

The brief mentions a field I did not describe. Is that a problem? It should not. Ask for it to be moved to open questions and marked "needs my confirmation" instead of guessed.

Complete reusable prompt

Test-and-Verify Brief for a Form-Filling Agent

You are helping me prepare a safe, reviewable test plan for an agent or browser helper that fills online forms. I am the human operator; the agent is the thing being tested. This prompt does not let the agent browse, submit, log in or send anything. It only produces a written brief I can follow and adapt.

What I will paste

I will paste all or part of the following. Do not fill gaps with guesses.

• Form name and purpose (for example, "school trip consent form").
• The field labels or a short transcript of what the form asks for.
• Which fields are optional versus required, if I know.
• The agent tool I plan to test, described in my own words.
• Whether a draft-save, preview or screenshot step is available to me manually.
• The most sensitive fields in the form, from my point of view.
• Any deadline or consequence of sending wrong information.

What to produce

Return a Markdown brief with these sections:

1. Scope and boundary

   • One sentence naming the form and the single test goal.
   • An explicit "This brief does not authorise" list: no submission, no payment, no account creation, no acceptance of terms, no email or message sending, no uploading of identity documents, unless I personally perform that step after review.

2. Sample data pack

   • A table with columns: Field label, Sample value, Why this value tests something, Real-data substitute to use later.
   • Use clearly fake values (for example, "Test Person") and mark each as SAMPLE.
   • Include at least one deliberately awkward case where relevant: a long name, a date near a month boundary, a phone number with country code, a special character in an address, or an optional field left blank.
   • Never include real addresses, account numbers, medical details or identity numbers in the sample column. If I pasted real data, replace it with a placeholder and note the substitution.

3. Checkpoint plan

   • Numbered checkpoints from "form opened" to "ready to submit".
   • For each checkpoint: what the agent should have done, what evidence I should capture manually (for example, a screenshot, a copied field dump, or a written note), and what would make me pause.
   • Mark any checkpoint that involves money, identity, consent, health or legal terms as HIGH-CAUTION.

4. Field-by-field verification table

   • Columns: Field, Expected sample value, What the agent put, Match? (Yes/No/Unclear), Action if not a match.
   • Leave the agent's output column blank for me to fill in during the test.

5. Stop rules

   • Hard stops: any submission, any request for a password or one-time code, any CAPTCHA bypass, any unexpected download or upload, any field where the agent's value cannot be traced to my sample pack, any HIGH-CAUTION field reached before I have reviewed it.
   • When a stop rule fires: stop, do not retry automatically, capture evidence, and report back to me.

6. Review checks before any live use

   • A short pass/fail checklist I run myself: all sample values replaced, all HIGH-CAUTION fields reviewed by a human, privacy notice read, submission step performed by me, confirmation saved.

7. Open questions and uncertainties

   • List anything you could not determine from what I pasted, and mark it "needs my confirmation" rather than inventing an answer.

Worked example (fictional, for shape only)

Input I might paste: "Form: community centre room booking request. Fields: name, email, phone, date, start time, end time, purpose, accessibility needs. Required: name, email, date. Sensitive: phone, accessibility needs. Tool: a browser agent that can type into fields. I can preview before sending."

Output shape:

• Scope: test whether the agent can complete the booking request fields from my sample pack without submitting.
• Sample pack row: Field "Date" | Sample "2026-02-28" | Tests end-of-month handling | Substitute my real requested date later.
• Checkpoint 3: agent has filled all fields, no submit pressed. Evidence: screenshot of the filled form. Pause if any field is blank or contains a value not in the sample pack.
• Stop rule: if the agent clicks "Confirm booking" or asks for a password, stop immediately.
• Verification row: Field "Phone" | Expected "+44 0000 000000" | Agent put: | Match?: | Action: retype manually if unclear.
• Review check: "All sample values replaced with real values" — Fail until I do it.

Checking steps for you

Before returning the brief, verify:

• Every sample value is labelled SAMPLE and contains no real personal data.
• Every HIGH-CAUTION field is named and has a human-review step.
• At least one stop rule covers submission and one covers credentials.
• Every unknown is listed under open questions instead of being filled in.
• The brief never says the agent has already performed an action.

Boundaries

• Do not claim the agent can or cannot do something; describe what to test and observe.
• Do not provide legal, medical or financial advice about the form's content.
• Do not tell me to disable security features or bypass checks.
• Keep the brief usable as a printed or on-screen checklist.

References and reuse

• ChatGPT release notes · Reviewed 2026-10-04

Original TokRepo prompt · CC BY 4.0. Reference documents retain their own rights.