Custom GPT Not Following Instructions? What to Check

When a custom GPT ignores its instructions, the most common reason is a conflict it cannot resolve: two rules that pull in opposite directions, or an uploaded file that contradicts what the instructions say. The model does not choose to ignore you; it chooses between you and something else in the setup, and picks the wrong one.

Custom GPTs are ChatGPT configurations that let you set standing instructions, upload knowledge files, and choose which capabilities the assistant uses. A tool like fenbs lets you store your tested prompts and record the lessons from each round of testing, so next time you have a starting point that actually works.

Look for conflicting instructions

Print your instruction set (or paste it into a plain text file) and read it as if you have never seen it before. Mark any rule that seems to contradict another. Common pairs: "reply briefly" alongside "explain your reasoning in full"; "stay strictly on topic" alongside "be helpful with related questions." When two rules conflict, the model must pick one. It may pick differently on different inputs, making the behavior look random.

Decide which rule wins when they collide. Make that explicit: "keep replies brief; if the user asks for an explanation, give no more than three sentences." Specificity beats generality, and an example beats either.

Check what your uploaded files say

If you added knowledge files, they can override or undercut your instructions without you realizing it. Say your instructions tell the GPT to recommend a specific workflow, but an uploaded document describes an older, different one. The model may treat the document as authoritative and follow it instead.

Review each uploaded file for statements that contradict your instructions. Remove or update any that do. If you need the file for reference but want the instructions to take precedence, say that explicitly: "when this file and the instructions conflict, follow the instructions."

Reduce instruction length

Long instruction sets are harder for a model to follow consistently. Every rule competes for attention, and rules near the bottom of a long list tend to get less weight. If your instructions run to many paragraphs, try a version with only the rules you have actually seen broken. If the GPT now follows them, add rules back one at a time, testing after each addition.

You do not need to explain every rule. "Reply in plain English" does not need a paragraph of rationale inside the instruction text. Save the rationale somewhere you can find it, but keep the instructions themselves as short as possible.

Add a concrete example

For any rule you have already simplified and still see broken, add a worked example to the instructions. Show a realistic input and exactly the output you want. The model uses examples more reliably than abstract rules because it can see the right answer rather than infer it.

Keep examples realistic. A made-up example like "User: Hello. Assistant: Hi there." does nothing. A real sample input from your actual use case, paired with the correct reply, gives the model something it can pattern-match against.

Start a new conversation to test

Always test a changed GPT in a fresh conversation. Existing conversations may carry context that masks the problem or creates new ones. Open a new chat, send a test input, and check the output before concluding the fix worked.

Reset with a minimal instruction set

If you have tried the steps above and the GPT is still unreliable, start with a very short instruction set: the single most important rule, one example, and nothing else. Test whether that works. If it does, add rules back one at a time, testing after each addition. The rule that breaks reliability is the one causing the problem.

This approach is slower than editing in place, but it isolates the cause rather than layering fixes on top of an undiagnosed problem. A GPT built up gradually from a working minimal state is more reliable than one edited repeatedly without knowing which change made a difference.

When you find the offending rule, decide whether you need it. If you do, rewrite it in more specific terms. If the behavior it was meant to produce happens reliably without it, the rule was redundant.

How to do this with fenbs

1. Create a free account and add a project for the work your GPT supports.
2. Save the current version of your instructions as a Skill in fenbs, with notes on what each rule is for.
3. Add a Lessons learned entry each time you find a conflict: what the conflict was and how you fixed it.
4. When you change the instructions, update the Skill so the current version is always recorded.
5. Before setting up a similar GPT, check the lessons from this one so you do not repeat the same conflicts.

For related problems, see custom GPT not updating if changes are not showing up, or ChatGPT custom instructions not working for issues with the personal instructions feature. Compare custom GPTs with other platforms in Claude Projects vs. custom GPTs.