Custom GPT Not Working? Checks and Recovery Steps

If a custom GPT is not working, first establish whether you cannot open it, cannot get a reply, or cannot complete one specific action. Check the signed-in workspace and any product transition notice before editing the GPT. Each symptom needs a different test, and repeatedly changing instructions can make the original failure harder to identify.

Keep the working brief and checked results outside the failing conversation. fenbs can hold that context for the next assistant you use, without depending on the custom GPT being available. It does not repair GPTs or automatically import their instructions, files, or chats.

Which quick checks narrow the problem?

1. Save the GPT link, exact error, time, and a safe copy of the request. Note whether the failure happens before or after you send it.
2. Confirm your account and workspace. Read any access or migration notice shown in the product.
3. Try a short ordinary conversation in ChatGPT, then a simple request in the GPT that does not need files or an outside service.
4. If basic replies work, test the failing capability separately with harmless sample information.
5. If you own the GPT, record its current setup before changing one relevant setting. Otherwise, send the evidence to its creator or workspace administrator.

Use one comparison at a time. An ordinary conversation working tells you that some ChatGPT functionality is available; it does not establish that every GPT capability is healthy. Likewise, a successful sample request does not prove that the original file or outside service is accessible.

Could the GPT have moved or lost access?

OpenAI documents an Enterprise transition in Moving your custom GPT workflows to plugins↗ (opens in a new tab). Ask the creator or administrator whether this workspace is affected and whether a replacement exists. Do not assume the same schedule or access rules apply to every account.

A migrated replacement starts private, so someone who could use the original may still need access to the plugin. Custom actions also need rebuilding separately. If your familiar GPT link stops being useful during that transition, ask for the tested replacement rather than assuming a browser reset can restore the old workflow.

For other access failures, have the owner compare the intended audience with your actual workspace membership. OpenAI’s GPT sharing and action policies↗ (opens in a new tab) explain that workspace controls affect sharing and permitted external domains. Record the message you receive; do not infer that a GPT was deleted simply because you cannot open it.

If work must continue elsewhere, write down the approved goal and remaining task before moving. fenbs AI context lets a connected assistant read that work record, while a person remains responsible for approving decisions and access.

Does the problem follow the browser or the conversation?

After copying any unsaved request, try the same small test in a fresh browser session with the same account. Check product notices for a service problem. If the fresh session behaves differently, review local site data or extensions before altering the GPT itself. Clear only what you understand and can restore.

Next, compare a new GPT conversation with the failing one, if that option remains available. Use a short prompt with a clear expected result, such as asking it to restate the task in two sentences. Leave out the large attachment and external action for this test.

Write the outcome narrowly: “new conversation answers; earlier conversation stalls on the attached report.” That gives the owner a useful starting point. “The GPT is broken” combines several possible causes and invites unnecessary rebuilding. If both tests fail, preserve the evidence and escalate rather than running an endless retry loop.

Is only an external action failing?

An action lets the GPT request information or do something through another service. OpenAI’s GPT Actions debugging guide↗ (opens in a new tab) recommends testing individual actions and inspecting their inputs and outputs. The creator should compare the requested operation, missing fields, and the outside service’s response.

For a sign-in failure, consult GPT Action authentication↗ (opens in a new tab). The application’s registered callback must match the intended GPT sign-in flow. Have the integration owner inspect the configuration and authentication logs; do not paste credentials or tokens into a conversation to bypass the failure.

• If the action never starts, record the request and which action should have been selected.
• If it starts and returns an error, preserve the safe error code and operation name for the integration owner.
• If the service succeeds but the answer is wrong, compare the returned result with what the GPT actually said.
• Before retrying an action that changes data, verify whether the earlier attempt already completed.

For example, an order lookup failing is different from an order submission with an uncertain outcome. Test the former with an approved sample. Check the destination system before repeating the latter, so an interrupted reply does not become a duplicate transaction.

What should you preserve before changing anything?

Make a small recovery record: the intended job, the last result you checked, the failing example, and what would count as success. If you are the creator, separately preserve the configuration and reference material you are authorized to retain. A summary of the job is not a backup of the GPT.

If replies arrive but the behavior is wrong, use the separate guide to custom GPTs not following instructions. A configuration that fails to save belongs with custom GPTs not updating. Keep this investigation focused on restoring usable access and identifying the failing component.

If you need a different workspace for the same job, compare Claude Projects and custom GPTs or ChatGPT Projects and custom GPTs. Move only approved material, then repeat the familiar test before relying on the replacement for real work.

How to do this with fenbs

1. Create a free account and add a project for the workflow the GPT supported.
2. Add a bug task with the failure, expected result, and a plan for the remaining checks. Keep credentials and private service logs out of the task.
3. Add concise AI context notes with reviewed findings and source links. Record a person’s approved requirements in Decisions and rules.
4. Use the integration guides to connect a supported assistant. Ask it to read context first and save Where we left off when the session ends.

fenbs is free for one person, no credit card. It is useful when the work must survive a change of assistant; it is not the right tool for restoring a deleted GPT or diagnosing a private API server. Those jobs still belong to the product owner or support team.

The AI at work hub has more workflows for continuing across tools. When the original problem is resolved, record the successful check and any remaining limitation so the next session starts from evidence.