Don't stop here
Hand-picked guides our readers explore right after this one.
Set up ChatGPT custom instructions for personalized responses
Read the guideExpert guide to Claude prompts with XML tags, artifacts, and complex reasoning
Read the guideAI prompts for content ideation, scriptwriting, repurposing, and audience growth strategies
Read the guideCustom GPTs break in a few recognizable ways: the GPT stops calling its Actions, it ignores the knowledge files you uploaded and answers from general knowledge instead (often confidently wrong), or knowledge files vanish after you save. Most of these trace back to two things: retrieval not firing (the GPT does not search its knowledge unless nudged), and regressions after OpenAI swaps the base model (a GPT tuned for one model can behave differently on the next). The good news is the fixes are quick, the highest-leverage one is a single instruction that forces the GPT to search its knowledge on every request. This guide covers Actions, knowledge files, and post-update regressions.
The GPT answers from general knowledge and ignores your uploaded files
It 'makes stuff up' that contradicts the documents in its knowledge
Actions stopped firing, the GPT no longer calls your API
Knowledge files disappear from the GPT after you upload and save
The GPT worked fine, then changed behavior after an OpenAI model update
It ignores instructions on the very first message of a session
A Custom GPT does not automatically search its knowledge files on every turn, it decides when to. If your instructions do not push it to, it often answers from general knowledge and skips the files entirely.
When OpenAI changes the default model behind ChatGPT (for example a GPT-5.x rollout), a GPT whose instructions were tuned for the old model can suddenly ignore actions or follow instructions differently. The GPT did not change, the model under it did.
A known bug where uploaded knowledge files save but then disappear, sometimes tied to the old GPT's stored configuration. Recreating the GPT fresh often sidesteps it.
If an Action's schema, auth, or endpoint changed, the GPT silently stops calling it. Re-saving the schema and re-checking auth usually restores it.
Some model builds under-apply the GPT's system instructions on the very first answer of a session, so the first response looks like the GPT is ignoring its setup.
When to try: First, for any 'ignores files / makes stuff up' problem
In the GPT's Instructions, add: 'Treat every user request as if it begins with: search your knowledge base for the answer. Always consult the uploaded knowledge files before answering, and if the answer is not in them, say so rather than guessing.' This single line fixes the majority of ignore-the-files cases.
When to try: When files are ignored or disappearing
Remove the knowledge files, re-add them, and save the GPT. Confirm they still appear after you reload the editor. Prefer clean text or well-structured PDFs, scanned/image PDFs are read unreliably. Keep individual files reasonably small.
When to try: When re-uploading and re-saving does not stick
If knowledge files keep vanishing or actions will not fire no matter what, build a new GPT and paste your instructions, then re-add files and actions. New GPTs often avoid the stored-config bug that haunts older ones.
When to try: When actions stopped firing
Open Configure, Actions, and re-save the schema. Re-check the authentication (API key or OAuth) and the server URL. A stale or slightly-off schema makes the GPT quietly stop calling the action. Test with a prompt that should trigger it.
When to try: When behavior changed right after a ChatGPT update
After an OpenAI base-model change, re-test your GPT and tighten the instructions: be explicit and imperative ('You MUST call the X action when...', 'ALWAYS cite the knowledge file'). Behavior that the old model inferred may need to be spelled out for the new one.
When to try: If only the first answer of a session is off
If the GPT ignores its setup on the first reply, add to Instructions: 'Before your first response in any session, silently re-read and apply all of the above instructions.' This nudges the model to apply the system prompt from message one.
Always include an explicit 'search your knowledge before answering' instruction
Keep knowledge as clean text or structured PDFs, avoid scanned/image files
Write instructions as explicit imperatives, do not rely on the model to infer intent
Re-test your Custom GPTs after every major ChatGPT model update
Keep a copy of your instructions and action schemas so you can rebuild the GPT quickly
Contact OpenAI via help.openai.com if knowledge files repeatedly disappear even on a freshly created GPT, or if a correctly configured Action returns errors that are clearly server-side. Include the GPT name, the exact behavior, and screenshots of the configuration. For 'ignores my files' the retrieval-forcing instruction fixes it faster than support can respond.
Because a GPT does not automatically search its knowledge on every turn. Add an instruction that forces it, for example 'treat every request as prefaced with: search your knowledge base', and tell it to say so when an answer is not in the files instead of guessing.
OpenAI periodically swaps the base model behind ChatGPT. A GPT whose instructions were tuned for the old model can start ignoring actions or following instructions differently. Re-test and tighten the instructions into explicit imperatives for the new model.
This is a known bug, often tied to an older GPT's stored configuration. Re-upload and save first, and if they still vanish, recreate the GPT from scratch and re-add the files, which usually sidesteps it.
Usually a stale or slightly-off Action schema, changed auth, or a changed endpoint. Open Configure, Actions, re-save the schema, re-check the API key or OAuth and the server URL, then test with a prompt that should trigger it.
Instruct it explicitly: always consult the uploaded files before answering, quote or reference the relevant part, and state clearly when the answer is not in the knowledge base rather than filling the gap with general knowledge.