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 guideCreate stunning images with ChatGPT DALL-E integration, viral photo trends, and professional visuals
Read the guideYou wrote clear custom instructions, and ChatGPT follows them for one message then drifts straight back to its default voice. This is one of the most reported ChatGPT annoyances of 2026, and it is rarely a bug. Custom instructions are not a hard filter, they are rewritten into a short system message at the start of each chat and then compete for attention with everything you type afterwards. Three things break them most often: saved memory holding an older preference that overrides what you just wrote, instructions written as vague aspirations rather than testable rules, and long chats or big file uploads that push the instruction block far back in context where the model weights it less. This guide covers each cause and the fixes that actually change output.
ChatGPT follows your instructions in the first reply, then reverts by the third or fourth
Tone rules (no em dashes, no bullet lists, keep it short) are obeyed inconsistently
Instructions appear saved in Settings but have no visible effect on answers
Behaviour differs between a new chat and an existing long thread
Custom GPT or Project instructions are ignored at the very start of a new chat
Editing and re-saving instructions appears to reset or truncate what you wrote
Memory entries saved weeks or months ago (for example a preference for formal tone from an old project) can outrank what you typed into custom instructions. OpenAI's own guidance is explicit that memory can override stated personality traits, so a stale saved fact quietly wins.
Custom instruction fields have a hard character cap. Long instruction sets get truncated at save time or squeezed into a compressed system message, so the rules at the bottom of your list are the first to be dropped.
Lines like 'be insightful and helpful' cannot be obeyed or violated in any measurable way, so they do nothing. Rules the model can be scored against ('never open a reply with the word Certainly') get followed far more reliably.
Asking for both thorough detail and extreme brevity, or both a warm conversational voice and strict formal register, leaves the model to pick. Contradictory rules cancel each other out and the default behaviour reasserts itself.
Instructions sit at the very top of the context window. After tens of thousands of tokens of conversation or a large uploaded document, they are far from the model's most recent attention and get weighted less than what you said two messages ago.
When to try: First, before rewriting a single line of your instructions.
Open Settings, then Personalization, then Manage memories. Read every entry. Delete anything that states a tone, format, or style preference that contradicts your current custom instructions, then start a brand new chat and retest. This single step resolves the majority of 'my instructions do nothing' reports.
When to try: When instructions are obeyed inconsistently rather than not at all.
Convert every aspiration into a rule that could be marked pass or fail. Replace 'write concisely' with 'keep replies under 200 words unless I ask for more'. Replace 'avoid corporate tone' with 'never use the words leverage, robust, or seamless'. If you cannot imagine grading a line, cut it.
When to try: For any rule phrased as a prohibition.
Negative-only instructions leave the model guessing. Instead of 'do not use em dashes', write 'do not use em dashes, use a comma or start a new sentence instead'. Instead of 'do not use bullet points', write 'do not use bullet points, write flowing paragraphs'. Compliance improves sharply when the substitute is spelled out.
When to try: When your instructions run to several paragraphs per field.
Count the characters in each field and keep each comfortably under 1,500. Rank your rules by how much you care, keep the top five to eight, and delete the rest. A short set that is fully applied beats a long set that gets truncated.
When to try: For work where format compliance really matters, such as client-facing drafts.
For the two or three rules you care about most, paste them again as your first message in a new chat. Instructions repeated in the conversation itself sit much closer to the model's attention than the settings block and survive far longer into a thread.
When to try: When adherence was fine early in a chat and degraded as it grew.
When a conversation passes roughly thirty exchanges or you have uploaded large documents, ask for a summary of where you are, open a new chat, paste the summary, and restate your key rules. Instruction adherence resets to full strength in the new thread.
When to try: When a Custom GPT or Project behaves correctly only after a message or two.
Custom GPT and Project instructions have been reported to be skipped at the very start of a new chat. Work around it by having the GPT's instructions end with a directive to restate its operating rules before answering the first message, which forces them into the visible conversation.
When to try: When you cannot tell whether instructions are reaching the model at all.
Clear all instructions and set one unmistakable rule, such as 'begin every reply with the word ACK'. Start a new chat and check. If that works, your instructions are being delivered and the problem is the content of your rules. If it does not, the problem is account-side and worth reporting to OpenAI.
Review saved memories monthly and delete stale style or tone preferences
Keep instruction sets short, ranked, and free of contradictions
Pair every prohibition with the behaviour you want instead
Restate your two or three most important rules at the top of any long or important thread
Contact OpenAI at help.openai.com if a single unmistakable test rule (like 'begin every reply with ACK') fails in a brand new chat with memory cleared, or if saved instructions repeatedly vanish from Settings after you save them. Those point to an account-side fault rather than the normal adherence drift described above.
Usually because a saved memory contradicts them, the rules are vague rather than testable, the instruction set is contradictory or over the roughly 1,500-character cap, or the chat has grown long enough that the instruction block is diluted in context. Clearing conflicting memories in Settings, Personalization, Manage memories fixes most cases.
Yes. Each custom instruction field is capped at about 1,500 characters. Longer sets get truncated or compressed, so rules at the end of a long list are the first to stop being applied. Keep each field well under the cap.
It can. OpenAI's guidance is that saved memories can override stated personality and tone preferences, so an old saved fact about how you like replies formatted may beat what you typed into custom instructions. Deleting the conflicting memory is the fix.
Instructions sit at the top of the context window. As a chat grows, or after you upload a large document, they are far from the model's most recent attention and get outweighed by recent messages. Restating the key rules mid-thread, or starting a fresh chat, restores adherence.
Project and Custom GPT instructions layer on top of your global ones, and users have reported them being skipped on the very first message of a new chat. Having the GPT restate its rules before answering the first message is a reliable workaround.