ChatGPT Custom Instructions Not Working? Find the Cause

When ChatGPT custom instructions are not working, check whether the instruction is present in the right scope, conflicts with other guidance, or describes behavior too vaguely to test. Then reproduce the failure with one simple request before changing the full instruction set.

For work rules that several assistants must follow, fenbs keeps the approved rule and its context outside personal settings. That gives you a current reference to check against, though you still need to test how each connected assistant uses it.

What should you check first?

1. Confirm the account, workspace, and surface where the failure occurred. Open the relevant settings and verify that the intended instruction is saved.
2. Read the actual task request and any project or reusable workflow instructions. Look for a different rule about the same behavior.
3. Replace the disputed instruction temporarily with one small, observable requirement and test it in a fresh conversation.
4. Restore other rules gradually. Keep the same request so you can identify which addition changes the result.
5. If the problem remains, save the request, output, instruction text, and setting details needed to reproduce it, without sharing sensitive information.

OpenAI’s personalization guide↗ (opens in a new tab) describes custom instructions as preferences across chats, with controls varying between web and desktop. Use the documentation for the surface you are testing; do not assume that two similar settings screens imply identical behavior.

A minimal test helps separate a setting problem from a difficult task. If “use three bullets” works on a simple summary but fails in a long request requiring a table, examine the competing requirements before concluding that personalization has stopped functioning.

Is the instruction in the place that needs it?

List where the relevant guidance lives: personal settings, the current request, a project, a selected reusable workflow, or a source file. Then check the location used by the failing task. A rule saved elsewhere may never have been supplied to that conversation.

For example, a team’s proposal rule might exist in one person’s chat while a colleague works in a separate project. The words are clear, but the second assistant has no established source for them. Copying a reminder into every request can hide this scope problem without fixing it.

OpenAI’s projects documentation↗ (opens in a new tab) says project instructions apply across that project’s chats. Test in the intended project and inspect its current guidance. Keep broader custom instruction examples separate from the project-specific rule you are trying to diagnose.

When the rule belongs to shared work, record the person’s choice in fenbs Decisions and rules, then ask the connected assistant to read project context before acting. This makes the reference easier to find when troubleshooting.

Do two reasonable rules contradict each other?

Conflicts are often ordinary preferences written at different times. “Always explain your reasoning in detail” clashes with “keep every answer under 100 words.” “Ask before making assumptions” can conflict with a task that explicitly asks for a first draft using stated assumptions.

Give each default an exception where the work needs one. Remove superseded versions instead of appending another “most important” sentence. Then rerun the same test, because a shorter and clearer rule is only an improvement if it produces the behavior you need.

Does the assistant lack facts or access?

An instruction can tell the assistant to consult the approved budget, but it cannot establish access to a missing document. Check whether the source is actually present, current, and readable through the available connection. Ask for the relevant source evidence rather than accepting “I checked” as proof.

Separate a missing fact from an ignored rule. If a request asks for the latest approved date but provides two undated versions, the useful behavior may be to ask which one controls the answer. Repeating “follow instructions exactly” does not resolve the ambiguity.

Memory also needs the right expectations. OpenAI’s memory guidance↗ (opens in a new tab) treats memories as recall rather than the sole source for required guidance. Keep a rule somewhere you can inspect and update, especially when people need to agree on what it says.

If a connection itself is failing, use the ChatGPT connector troubleshooting guide. A missing connection and an instruction conflict need different fixes; changing wording will not repair access.

What can instructions never guarantee?

Treat an instruction as guidance to verify, not a substitute for a control that must reliably prevent an action. A sentence asking for approval is useful, but sensitive workflows also need appropriate permissions, available-tool settings, and a person reviewing the result.

Do not judge success by asking the assistant to repeat your rules. Repeating an instruction and applying it are different tasks. Give it a case that requires the behavior: an ambiguous date, a source with a contradiction, or a request that should remain a draft.

If the assistant still fails a clear, simple case, keep the reproduction and follow the product’s support route. Record what you observed rather than claiming to know the internal cause. A useful report says which instruction was missed and supplies the smallest example that demonstrates it.

How to keep instruction fixes with fenbs

1. Create a free account and add the project affected by the instruction failure.
2. Create a task that names the observed behavior and the result you expected. Put the reproduction and checks in its plan.
3. Save the approved instruction in Skills and prompts and record the person’s lasting requirement in Decisions and rules.
4. Run the normal and difficult test cases in your assistant, record the results on the task, and add a lesson if the failure should change future work.
5. Save Where we left off with the working version and any unresolved failures so another session can continue the diagnosis.

fenbs does not change ChatGPT’s personalization settings or guarantee that a model obeys every instruction. If you only need to adjust a private style preference, fix it inside ChatGPT. Use fenbs when the requirement belongs to ongoing work that people and connected assistants need to share.

For failures beyond instructions, use the separate guides to ChatGPT availability, memory, projects, deep research, or scheduled tasks.

The AI prompts hub covers reusable guidance. Keep your successful test with the rule so the next troubleshooting session starts from evidence.