ChatGPT ignores your instructions in a custom GPT or Project: why it happens and how to fix it

1 hour ago

8 min read

Last verified: November 4, 2026

You spend time writing careful instructions for a custom GPT or a ChatGPT Project, and then the very next answer behaves as if they were never there: the wrong tone, the wrong format, a banned word, an old fact from a previous chat. The complaint fills the official OpenAI Developer Community, with threads that run from early 2025 to this autumn, and Google mostly sends people to those threads without a clear answer. The cause is rarely a single bug, because ChatGPT combines your GPT or Project instructions with OpenAI's own rules, your personal custom instructions, your memory and the earlier conversation before it writes a reply.

The most common fix is shorter instructions with no conflicts, plus a fresh chat. OpenAI's own prompting guidance says contradictory or vague instructions hurt its recent models more than older ones, and its instruction hierarchy means a GPT or Project can never override rules set at a higher level. When that is not enough, check memory and past-chat references, move rules out of uploaded files into the instructions themselves, and report reproducible failures with an example. Owners of custom GPTs also face a deadline: OpenAI is retiring custom GPTs in favour of plugins, and its migration guide warns that a migrated plugin may respond differently from the original GPT.

··········

WHAT IGNORED INSTRUCTIONS LOOK LIKE IN CHATGPT GPTS AND PROJECTS.

The symptom takes a few recognisable forms, and naming yours correctly saves time, because each one points to a different cause. Some users see the instructions ignored from the first reply of every new chat; others see a good start that drifts after a long session; others find that a Project keeps repeating details from an earlier conversation even after the instructions were changed. A separate group reports GPTs that answer like plain ChatGPT, as if no configuration had loaded at all.

........

What you see	Where it is reported	Most likely area to check
First reply in a new chat ignores the instructions	GPTs and Projects, reported with GPT-5.5 in June 2026	Report it with an example (see below)
Rules followed at first, dropped after a long session	Custom GPTs, desktop app	Long conversation, start a new chat
Project repeats facts from an older chat	Projects	Chat history and memory references
GPT acts like default ChatGPT, no persona at all	Custom GPTs, web and iOS, late 2025	Configuration not applied, report it
Some rules followed, others never	GPTs and Projects	Conflicting or overlong instructions
Rules written in an uploaded file are ignored	Custom GPTs with knowledge files	Move rules into the Instructions box

........

··········

WHY CHATGPT IGNORES GPT AND PROJECT INSTRUCTIONS: CAUSES, CHECKS AND FIXES.

OpenAI describes an instruction hierarchy for its models in which system instructions rank above developer instructions, which rank above user messages, which rank above tool outputs. Higher-priority instructions are more trusted, and the company says a lower-priority instruction that tries to override a higher-priority one should be declined. Inside ChatGPT, that means a GPT or Project cannot switch off OpenAI's own system rules or usage policies, however firmly the instruction is worded. The causes below cover what remains under your control.

........

Cause	How to check	Fix
Instructions contradict each other	Read them for rules that cannot both be true	Remove or rank the conflicting rules
Instructions are too long or vague	Count the rules and look for soft wording	Keep the few rules that matter, state them plainly
Instruction asks for something higher rules forbid	The model refuses or answers in its standard way	Rewrite the request within OpenAI's policies
Memory or earlier chats pull in old context	Ask what it remembers, review saved memories	Delete the memory or turn off chat-history references
Very long conversation	Problem starts late in a long session	Start a new chat in the same GPT or Project
Rules stored only in a knowledge file	The rule lives in an uploaded document	Put behaviour rules in the Instructions box
Configuration not applied (reported bug)	GPT behaves like default ChatGPT from message one	Collect an example and report it to OpenAI

........

··········

HOW TO FIX CONFLICTING OR OVERLONG CHATGPT INSTRUCTIONS.

OpenAI's GPT-5 prompting guide says the model follows prompt instructions with "surgical precision", and for the same reason "poorly-constructed prompts containing contradictory or vague instructions can be more damaging" than with earlier models, because the model spends reasoning effort trying to reconcile the contradictions instead of picking one. The company's troubleshooting guide adds that, in the cases its team investigated, erratic behaviour "stemmed from a contradiction between separate sections of the prompt". A GPT built two years ago for an older model and edited piece by piece is a typical place for such contradictions to hide.

1. Copy the full instructions into a plain document and list every rule as a single line.

2. Mark pairs that clash, for example "always answer in one paragraph" next to "always include a table", or "never ask questions" next to "ask for missing details".

3. Decide which rule wins in each pair and delete the other, or write the priority explicitly ("if the user asks for a table, the one-paragraph rule does not apply").

4. Cut rules that only restate the obvious. OpenAI's ChatGPT prompting page recommends focusing on "the one or two boundaries that matter most".

5. In a custom GPT, stay within the Instructions box limit: builders on the official forum report a hard cap of 8,000 characters, beyond which the editor will not save.

6. Save the GPT or Project, open a new chat, and test the rule that failed with the same message as before.

One user on the official forum reported that asking ChatGPT to rewrite the instructions in its own words, then explaining them back, made a GPT "work much more reliably". It is a single report, so treat it as something to try after the clean-up above, and check the rewritten text for new contradictions before saving.

··········

HOW TO STOP CHATGPT MEMORY AND OLD PROJECT CHATS FROM OVERRIDING INSTRUCTIONS.

A Project is built to carry context forward, and that is also why it can keep using outdated information. OpenAI's own Projects guide explains that chats inside a project can reference other conversations within the same project, and Business and Enterprise users can create a project with project-only memory, where saved memories from outside the project are not referenced. A user on the official forum described exactly this side effect: after the first chat in a project mentioned one servo motor model, new chats kept citing it even though the project instructions named a different one.

1. Ask ChatGPT inside the GPT or Project what it currently remembers about the topic, so you can see which old facts it is relying on.

2. Open your personal settings and review saved memories; delete any entry that contradicts the instructions (an old preferred format, an outdated product name).

3. If the problem persists, turn off the options named "Reference saved memories" and "Reference chat history" in your personal settings and test again.

4. In a Project, delete or archive the older chat that contains the outdated detail, then start a new chat.

5. On Business and Enterprise, consider recreating sensitive or long-running work as a project with project-only memory, which OpenAI says is available when memory is enabled for the workspace.

Your personal custom instructions, set under Settings and Personalization, apply across chats, and the personality chosen under Base style and tone changes style and tone without changing capabilities, according to OpenAI's personalization page. If your personal instructions ask for a style that a GPT or Project asks you to drop, remove the clash at its source. Our explainer on how ChatGPT memory and past chats work covers the memory settings in more detail.

··········

WHEN CHATGPT IGNORES INSTRUCTIONS IN THE FIRST REPLY OR IN LONG CHATS.

Two patterns appear in the forum reports that no settings change has been shown to fix. In June 2026 two forum threads reported that GPT-5.5 ignored custom GPT and Project instructions at the very beginning of every new conversation, and one stressed that the instructions had been checked for contradictions. An OpenAI Support representative replied in September asking whether the problem continued and requested "a recent example with the saved instructions, your first prompt, and the response that didn't follow them". OpenAI has not published a cause. In late 2025 other users reported custom GPTs that loaded with no persona at all on the web, and one user later reported the same on iOS; clearing the cache, private windows and other browsers did not help them.

The second pattern is drift in long sessions: rules respected at first and dropped after an extended conversation, reported by GPT creators on the official forum. OpenAI has not documented why this happens inside GPTs. Our guide to ChatGPT context windows and long chats explains how much of a conversation the model can take into account.

1. Start a new chat in the same GPT or Project and resend the request; if it works there, the long conversation was the problem.

2. If the first reply of a new chat fails repeatedly, save the instructions, your first message and the reply that ignored them.

3. Send that example to OpenAI Support, which is the evidence its staff asked for on the forum.

4. As a stopgap, users report pasting the key rule into the first message; one user said this works, another said reminders did not help.

··········

CUSTOM GPTS BECOME PLUGINS: WHAT CHANGES FOR YOUR CHATGPT INSTRUCTIONS.

OpenAI's documentation states that it is "transitioning custom GPTs to plugins". When retirement takes effect, custom GPTs stop running and leave the GPT directory, and creation of new GPTs is ending. The standard retirement date communicated to workspaces is December 11, 2026, and affected Enterprise workspaces with an approved deferral can keep them until February 11, 2027. The migration changes the way your instructions are applied, so a GPT that worked well may start ignoring rules after it becomes a plugin.

........

GPT component	What happens in a plugin	What to check after migrating
Instructions	Become a skill	Test the rules that matter most
Knowledge files	Become skill reference files	Retrieval may differ, test answers that cite them
Selected model	Does not carry over, Enterprise defaults apply	Answers may change style and accuracy
Tools (web search, data analysis)	Not guaranteed to match	Re-enable and test each tool
Custom actions	Do not transfer	Rebuild with an app or a custom MCP server
Conversation starters	Do not transfer one-to-one	Recreate the starters you need

........

OpenAI also notes that "an installed plugin won't run on every request", because skill selection depends on the task. A rule that a GPT applied to every message can therefore appear to be ignored after migration simply because the plugin was not selected for that request. Test the migrated plugin with the same prompts that the GPT handled well, and keep the original GPT until you are satisfied with the result. Our overview of custom GPTs for business teams describes how GPT instructions were designed to work.

··········

WHAT OPENAI DOCUMENTS AND WHAT USERS ONLY REPORT ABOUT CHATGPT INSTRUCTIONS.

........

Point	Status
Higher-priority instructions override GPT and Project instructions	Documented by OpenAI (instruction hierarchy)
Contradictory or vague instructions reduce adherence	Documented in OpenAI's prompting guides
Project chats can reference other chats in the same project	Documented in OpenAI's Projects guide
Project-only memory on Business and Enterprise	Documented in OpenAI's Projects guide
Migrated plugins may respond differently from the GPT	Documented in OpenAI's migration guide
8,000-character limit on GPT instructions	Reported by builders on the official forum
First reply of new chats ignoring instructions (GPT-5.5)	Reported by users, no published cause
GPTs loading with no instructions on web and iOS	Reported by users in late 2025, no confirmed fix
Drift in long sessions	Reported by GPT creators, not documented

........

Usage caps are a separate matter and do not by themselves make ChatGPT ignore instructions; our guide to ChatGPT usage limits by plan covers them.

··········

CHATGPT GPT AND PROJECT INSTRUCTIONS: WHAT TO CHECK FIRST.

Most cases of ignored instructions in ChatGPT trace back to instructions that disagree with each other, with a higher-level rule, or with something ChatGPT remembers from earlier chats. Rewriting the instructions as a short list of non-conflicting rules, clearing stale memories and starting a fresh chat resolves the documented causes. The remaining cases, such as the first-reply failures reported with GPT-5.5 and GPTs loading with no configuration, have no published explanation, and the only route OpenAI staff have offered is a support report with a concrete example.

GPT owners have a firm date to plan around. With custom GPTs scheduled to stop running on December 11, 2026 for most workspaces, the instructions you fix today will be converted into a plugin skill, with a different model and task-based activation, so the same tests need to be run again after migration.

··········

·····

DATA STUDIOS

·····

datastudios.org