How Do I Fix ChatGPT Custom Instructions Not Being Followed?

Issue type: AI behavior — ChatGPT’s custom instructions (set in Settings > Personalization > Custom instructions) are ignored, partially followed, or inconsistently applied across conversations

SEO focus: If you are looking for chatgpt custom instructions not being followed, this guide starts with the fastest diagnosis, then moves to the exact fix that matches the symptom.

Short answer: Custom instructions in ChatGPT are not commands — they’re context added to the beginning of every conversation. When instructions are ignored, it’s almost always because a long conversation pushed them toward the edge of the context window, the instruction wording is too vague to influence behavior, or the specific request in the conversation contradicts the instruction and the request wins. Rewriting instructions as specific constraints rather than general preferences dramatically improves adherence.

ChatGPT Custom Instructions Not Being Followed: Quick Diagnosis

Custom instruction failures have four distinct patterns — each with a different root cause:

Instructions are completely ignored from the first message: Either custom instructions aren’t enabled, or the instruction text contains something that conflicts with OpenAI’s safety guidelines, causing the system to deprioritize the entire instruction block. Check Settings > Personalization > Custom instructions and confirm both boxes are enabled and the text is saved.

Instructions are followed at the start but stop being followed after many messages: Classic context window fade. As conversations grow longer, the custom instructions occupy an increasingly small proportion of the total context. Models weight recent context more heavily than distant context — by message 30, the instruction from the beginning of the conversation has relatively little influence compared to the momentum of the conversation itself.

Instructions are followed for some requests but not others: The instruction is too general and the model’s interpretation varies by request type. “Be concise” means different things for a one-sentence question versus a request to summarize a 10-page document. Specific, measurable instructions adhere more reliably than stylistic preferences.

Instructions worked before a ChatGPT update and stopped working: OpenAI periodically adjusts how the system prompt (which includes custom instructions) is weighted against the conversation. A model update can change this weighting, which sometimes requires rewriting instructions to be effective with the new model version. If ChatGPT’s general consistency changed around the same time, a model update affected both behaviors.

What to do first

1. Confirm custom instructions are enabled and saved. Click your profile picture at the bottom-left of ChatGPT > Settings > Personalization > Custom instructions. Check that both toggles at the bottom are on: “Enable for new chats” must be active. If the text field is empty or the toggles are off, instructions won’t apply. After making changes, start a brand new conversation — custom instructions don’t apply retroactively to existing chats.
2. Test in a fresh conversation immediately after confirming instructions are saved. Long conversations dilute instruction influence due to context window mechanics. If your instructions are followed in a new short conversation but ignored in the same 40-message conversation you’ve been running, context window fade is the cause. Start a new conversation for the task where you need the instructions followed.
3. Rewrite instructions as specific constraints rather than general preferences. Compare these two approaches:
   Vague (low adherence): “Be concise and professional.”
   Specific (high adherence): “Limit responses to 3 sentences unless I explicitly ask for more detail. Never use bullet points. Use formal English without contractions.”
   Specific, measurable constraints give the model clear pass/fail criteria it can evaluate against. Vague stylistic preferences are interpreted subjectively and inconsistently. The more precisely you define what compliance looks like, the more reliably the model follows the instruction.
4. Remove conflicting instructions. Instructions that contradict each other cause the model to resolve the conflict in unpredictable ways. For example: “Always provide multiple perspectives” combined with “Be concise” creates a contradiction the model resolves differently on each turn. Review your custom instructions for internal conflicts and pick one behavior when two preferences can’t coexist.
5. Use the “What I want ChatGPT to know about me” field strategically. This field is for facts about you (profession, expertise level, goals). The “How would you like ChatGPT to respond?” field is for behavioral instructions. Mixing factual context into behavioral instructions dilutes both — the model processes them together and each competes for influence. Keep them in their intended fields for maximum effectiveness.
6. Reinforce critical instructions in the conversation itself when needed. For a conversation where following a specific format is essential, start your first message with a brief reminder: “Remember: respond in bullet points only, max 5 points per response.” This places the instruction close to the current context position where it has full influence, rather than relying solely on the system-level instruction that may be far back in the context. This approach is especially useful for long ongoing conversations where context fades — brief mid-conversation reminders restore adherence without starting a new chat.

Common mistake

Writing custom instructions as a personality profile or wishlist rather than behavioral constraints. Instructions like “I prefer creative, thoughtful responses that show genuine engagement with my questions” describe a desired feeling but give the model no specific behavioral target. The model’s interpretation of “creative” and “thoughtful” varies with every request. Effective instructions tell the model what to do and what not to do in measurable terms — length limits, format requirements, topics to avoid, response structure. Instructions written as feelings are suggestions; instructions written as rules are constraints. The model treats them very differently.

Best next step

If specific, measurable instructions still aren’t being consistently followed, test whether the issue is the instruction content or the system. Create a minimal test instruction: “Always start every response with the word CONFIRMED.” Save it, start a new conversation, and ask a simple question. If ChatGPT doesn’t start with “CONFIRMED,” custom instructions aren’t being applied at all in your account — this points to a settings or account issue rather than an instruction-writing problem. If it does start with “CONFIRMED,” custom instructions are working but your behavioral instructions are too loosely worded to enforce. This binary test is the fastest way to separate “instructions aren’t loading” from “instructions aren’t clear enough.”

Official reference: If the issue appears across devices or networks, check the official OpenAI Status page before reinstalling the app or changing account settings.

Quick Q&A

I asked ChatGPT to always respond in Spanish via custom instructions, but it keeps responding in English. Why?

Language instructions are some of the hardest to enforce through custom instructions because the model defaults strongly to responding in the same language as the user’s message. If you write in English, ChatGPT will typically respond in English even if told otherwise in custom instructions. The most reliable way to enforce Spanish responses is to write your messages in Spanish — ChatGPT almost always mirrors the input language. Alternatively, prefix your instruction with a higher-priority signal: “IMPORTANT: Always respond in Spanish regardless of the language of my messages.” The word “IMPORTANT” and the “regardless of” clause both help override the default language-matching behavior.

Do custom instructions apply to all models in ChatGPT, including GPT-4o and o1?

Custom instructions apply to GPT-4o by default. Behavior with other models (o1, o3) varies — some reasoning models have different system prompt handling that can reduce the impact of custom instructions, as these models prioritize internal reasoning chains over stylistic constraints. If you switch between models frequently and notice inconsistency in instruction adherence, the model change is likely the cause. Stick to one model for work where custom instruction adherence matters and test the specific model’s behavior with your instruction set.

Custom instructions disappeared after I logged out and back in. Where did they go?

Custom instructions are stored on OpenAI’s servers tied to your account, so they should persist across logouts and device changes. If they disappeared, the most likely cause is that you were in an incognito window when you set them (session cleared on close), you have multiple ChatGPT accounts and you’re now logged into a different one, or a rare sync error cleared them. Go to Settings > Personalization > Custom instructions and re-enter them. For safekeeping, copy your custom instructions to a notes app so you can quickly restore them if they’re cleared again.