This is maintenance for an existing GPT. Creating a GPT has separate eligibility and workspace restrictions; a subscription purchase is not a reliable repair for a wrong answer.
Reproduce the answer with a small reference fixture
Consider a fictional equipment-loan assistant. Its old handbook says returns go to the front desk. A new approved reference says returns go to the equipment counter, except weekend returns, which require prior arrangement.
Use a permitted test file with plain text:
Equipment returns — practice reference
Status: approved practice version B
Weekday returns: equipment counter.
Weekend returns: prior arrangement required.
Overdue fee: not specified in this reference.
Ask the GPT where a weekday return goes, what applies at weekends, and what the overdue fee is. The expected result is the counter, the arrangement condition, and an acknowledgment that the fee is unspecified. A plausible fee is a failed answer even if the destination is correct.
The fixture is an authored test case, not a real lending policy or an observed ChatGPT result.
Separate missing information from retrieval trouble
Open the uploaded source and find the sentence yourself. If it exists only in a picture of a notice, add clear text describing the relevant rule. OpenAI recommends text-forward reference files and cautions against relying on images inside files. Source-format checks
If two files conflict, decide which is authoritative before requesting another answer. Keep only the approved reference needed for this task, or clearly label the historical file and explain its limited use. Do not ask the GPT to infer approval from a filename that merely says “new.”
A narrower test can reveal what still fails:
Use "Equipment returns — practice reference", version B.
Answer these separately:
1. Where do weekday returns go?
2. What condition applies to weekend returns?
3. Does the reference specify an overdue fee?
For each answer, name the supporting section and give a short
supporting passage. When a fact is absent, say it is not specified.
Do not use general equipment-rental policies to fill the gap.
Check each passage in the actual source. A quotation and a filename in the answer still need comparison with the uploaded text.
Repair the instruction only when the test needs it
Put a short rule in Instructions requiring reference-supported answers and explicit acknowledgment of missing facts. Remove contradictory rules such as “always give an exact amount” when the source may omit the amount. Keep the destination and weekend policy in the reference file so future policy revisions do not require changing the same fact in two places.
Repeat the original three questions after each meaningful change. Save the actual responses alongside the fixture and the source version; record which check failed instead of accepting a general “looks better.”
Keep the test when the workflow moves
OpenAI is transitioning custom GPTs to plugins. Review the notice for the GPT's account or workspace and test the replacement before switching. Migration can copy knowledge into reference files, but equivalent answers require checking. Official migration guidance
Run the counter, weekend and missing-fee questions against the replacement too. Keep using the approved lending source as the policy record. The useful result is a correctly sourced answer, including an honest limit when the policy does not say.