ChatGPT / ChatGPT Research & Iteration
ChatGPT: keep an answer inside a supplied evidence set
Provide two short approved policy excerpts and label them A and B. Include one question the excerpts do not answer.
Environment and inputs: ChatGPT user guide · Feature and source access depend on the active account, app and organization settings. Optional file or app steps have a text-excerpt fallback. No fixed model version, price or usage quota is assumed.
What you will make
A practical ChatGPT guide to keep an answer inside a supplied evidence set. Use the sample input and prompt, then check the result. Use a fictional small-office IT team as the practice setting. Keep the approved facts in a clearly labeled source block.
Prepare the source and scope
Provide two short approved policy excerpts and label them A and B. Include one question the excerpts do not answer. Open the ChatGPT app or website and start a focused conversation for this exercise.
Try this prompt
In ChatGPT, provide the prepared input, then send the following request. For real work, replace fictional details with approved facts.
Answer the questions using only A and B. Reference paragraph labels. For anything not established by these excerpts, say Not established and list the needed evidence.Review the result
The unanswered question should remain unresolved, even if a plausible general answer exists. Record each correction and its supporting source.
If the answer is wrong or incomplete
General knowledge can leak into a source-limited task. Require support for every policy claim. For recurring work, keep the active source list with the project brief; for a one-off task, supply just the relevant excerpt.
Practice and completion check
Repeat with one incomplete input. Check the answer against the review criteria, remove unsupported claims and save the prompt with its source version. This workflow has not been verified in a customer deployment.