A reusable field template
A prompt testing kit with the context built in.
Record what you asked, what answered, which page was cited, and whether the claim was supported. Build a useful observation history without disguising uncertainty.
UTF-8 text file · editable in any text editor · no account required.
Choose the questions deliberately
Start with a defined audience and topic. Include questions about definitions, evaluation criteria, requirements, implementation, and relevant limitations. Keep branded verification separate from unbranded discovery.
Replace bracketed terms with real, non-sensitive context before using the prompts. Do not ask a question that instructs the assistant to mention your brand when you are trying to observe unbranded discovery.
Example prompt families
Definition: What is [concept], and when is it useful?
Evaluation: What criteria matter when choosing [category] for [use case]?
Comparison: How do [approach A] and [approach B] differ for [task]?
Requirements: What prerequisites apply before using [process]?
Branded verification: Does [product] support [capability]?
Record the conditions
For each observation, preserve the exact prompt, assistant, mode, language, timestamp, and relevant context. Keep the response where permitted and distinguish a valid answer from a failed or unusable request.
Record raw citation URLs before normalizing them for a report. Review whether the page supports the statement, including the conditions and version. Use supported, unsupported, ambiguous, or not reviewed as clear working labels.
Use a repeatable counting rule
For a simple pilot, count an answer once if it contains at least one qualifying citation to the target domain. Divide by the valid answers in the same stated sample. Report failures separately and keep the denominator visible.
Finish with a reviewable next action
A useful observation connects to an audience question and, where appropriate, an affected page. Record the finding, the evidence, the proposed action, the owner, and the acceptance check. A factual contradiction needs review; a single unexplained variation may need another observation rather than an immediate rewrite.
Keep the initial question set stable during a comparison period and annotate changes. Do not interpret the sample as all activity on an AI platform. The kit does not automatically query services or connect to accounts; it is a structure for your own authorized research and documentation.
Use the citation tracking guide for source-review examples and the metrics guide for reporting definitions.