Attach Knowledge Files for Grounded Answers

A grounding step for Custom GPTs: upload the authoritative documents the GPT should answer from — policies, specs, FAQs — and instruct it to rely on them rather than its general knowledge.

Without source files, a Custom GPT answers from generic training, which is fine until it confidently states something wrong for your specific context. Knowledge files anchor it to your actual material, so answers reflect your policies and facts — and you can point it to say 'not covered' instead of inventing, which is what makes a support or reference GPT trustworthy.

Any Custom GPT that should answer about specific, stable material: company policies, product docs, a style guide, a curated FAQ.

General-purpose assistants meant to draw on broad knowledge, or material that changes so often the uploaded files would be stale — those need a live source, not attachments.

For this Custom GPT, I've uploaded <the documents>. In the instructions, add: answer from these files first; quote or reference the relevant part; if something isn't covered by the files, say so plainly rather than guessing from general knowledge.

An HR-policy GPT grounded in the actual handbook answers 'how many sick days do I get?' from the document — and says 'that's not in the handbook, check with HR' instead of inventing a number.

• Relying on general training for questions that need your specific facts.
• Uploading files but not instructing the GPT to prefer them.
• Leaving stale documents attached, so it answers from outdated policy.