You set “Always answer in bullet points, never use marketing fluff, write at a 10th-grade level” in Custom Instructions weeks ago, and for a while ChatGPT obeyed. Now you ask a question and get back a flowery three-paragraph essay, no bullets in sight. Toggling the panel off and on does not help. The instructions are still saved; the model just is not following them.
Fastest fix: open a brand-new chat outside any Project and ask a question that should clearly trigger a rule. If the new chat obeys, your problem is scope (a Project or a long chat), not the saved instructions. If even a fresh chat ignores them, the cause is upstream: a conflicting Memory entry, a Personality preset, or instructions that are too long, too vague, or simply toggled off.
As of June 2026, Custom Instructions are one of three personalization layers ChatGPT stacks on every chat: the Personality preset, Custom Instructions, and saved Memory. They live under Settings -> Personalization (some clients label the entry Customize ChatGPT). When instructions get ignored, the cause is almost always one of those three colliding, or scope, overflow, or a model switch, not the panel itself being broken.
Which bucket are you in
| Symptom | Most likely cause | Jump to |
|---|---|---|
| Fresh chat obeys, this one does not, and it sits inside a Project | Project instructions override account ones | Cause 1 |
| Even brand-new chats ignore one specific rule | A saved Memory contradicts it | Cause 2 |
| Tone is wrong (too chatty / too blunt) but format is fine | Personality preset fighting your tone rules | Cause 3 |
| Started obeying, drifted after ~30 turns or a file upload | Context overflow pushed instructions out | Cause 4 |
| Replies got shorter / looser right after a “limit” notice | Silent fallback to GPT-5.4 mini | Cause 5 |
| Some rules followed, others never | Instructions too long, contradictory, or vague | Cause 6, 7 |
| Nothing applies at all, fields look populated | ”Enable for new chats” toggle is off | Cause 8 |
Common causes
Ordered by hit rate, highest first.
1. Project instructions overrode them
If the chat is inside a Project, the Project’s instructions take precedence and the model follows those instead of your account-level Custom Instructions.
How to judge: Look at the left sidebar. Is this chat nested under a Project? If yes, open the Project, click the three-dot menu, then Instructions (older clients call it Configure), and read them. They are likely contradicting yours.
2. A saved Memory contradicts the instruction
This is the most common cause that survives a fresh chat. Memory persists facts and preferences across all chats, and a stale Memory like “User prefers detailed prose explanations” will quietly outweigh a Custom Instruction that says “use bullet points.” If even a new chat ignores one specific rule, suspect Memory first.
How to judge: Open Settings -> Personalization -> Memory -> Manage memories. Skim the list for any entry that contradicts the rule being ignored, and delete it.
3. A Personality preset is fighting your tone rules
As of June 2026 ChatGPT ships preset personalities (Default, Friendly, Efficient/Professional, Candid, Quirky, Cynical/Robot, Nerdy). The preset shapes tone and can override soft tone rules in your instructions, while leaving hard format rules alone. Classic signature: bullets are correct but the voice is wrong.
How to judge: Open Settings -> Personalization. If the personality is set to anything other than Default, set it back to Default and re-test, or fold the tone you want into Custom Instructions instead.
4. Context overflow pushed instructions out of attention
Custom Instructions are prepended to the system prompt, but on a long chat the model weights recent tokens more heavily, so the instructions fade. Uploading a PDF or running a tool inserts large chunks of text above your latest message and drowns them out even faster.
How to judge: Did the chat obey at first and only drift after ~30 turns, or right after a file upload or a code-interpreter run? That is the signature of context pressure.
5. The model was silently switched mid-conversation
When you hit a usage cap, paid accounts get rerouted to a smaller model that follows long system prompts less strictly. As of June 2026 the picker shows GPT-5.5 (Instant / Thinking / Pro), and the silent fallback is GPT-5.4 mini, which does not appear in the picker at all. On Plus/Go you get roughly 160 messages every 3 hours before the fallback kicks in.
How to judge: Replies suddenly getting shorter or looser right after a limit notice is the tell. The fallback is not labeled in the picker, so judge by behavior and timing, not the model name shown.
6. Custom Instructions are too long or contradictory
Each field caps at 1,500 characters (3,000 total across both fields, as of June 2026). Pack them full, or mix rules like “be concise” and “always include 3 examples”, and the model picks one or follows neither consistently.
How to judge: Open Settings -> Personalization -> Custom Instructions and read both fields. If either is near the 1,500-character cap or contains opposing rules, that is the issue.
7. The model interprets a soft rule differently than you do
A rule like “use British English” is followed; a rule like “be more direct” is fuzzier, and the model often regresses to its default tone under load. Vague rules get ignored first.
How to judge: Replace one soft rule with a hard, testable one (“end every reply with the word DONE”) and watch it. If hard rules stick and soft ones do not, the instructions just need rewriting.
8. The instructions toggle is actually off
The “Enable for new chats” toggle can get flipped off, especially after a settings sync. The fields look populated but the toggle below them is off.
How to judge: Settings -> Personalization -> Custom Instructions, confirm Enable for new chats is on.
Before you start
- Have your current Custom Instructions text in front of you so you can edit.
- Decide whether to keep account-level instructions, Project-level, or both with clear roles.
- Pick one or two rules that matter most, so you can test compliance after the fix.
Information to collect
- Exact text of both Custom Instructions fields (
What would you like ChatGPT to know about you?andHow would you like ChatGPT to respond?). - Whether the current chat is inside a Project, and the Project instructions if so.
- The personality preset currently set, and a quick skim of saved Memory.
- The model shown in the picker for the failing chat, and whether you recently saw a limit notice.
- Approximate turn count and attachment count in the chat.
- Plan tier (Free / Go / Plus / Pro).
Step-by-step fix
Step 1: Confirm Custom Instructions are enabled
Go to Settings -> Personalization -> Custom Instructions. Both fields should have text and Enable for new chats should be on. Save. Toggling off and on triggers a fresh apply.
Step 2: Test with a brand-new chat
Start a fresh chat outside any Project and ask a question that should clearly trigger a rule (for example, “explain photosynthesis”, expecting bullets if you specified bullets). A fresh chat is the cleanest test of account-level instructions and isolates scope from content.
Step 3: Clear conflicting Memory
Open Settings -> Personalization -> Memory -> Manage memories. Delete any saved entry that contradicts a rule being ignored (for example a “prefers long prose answers” memory when you want bullets). If you cannot find the offender, turn Memory off temporarily and re-test; if behavior corrects, a memory was the cause.
Step 4: Reset the Personality preset
In Settings -> Personalization, set the personality back to Default if it is anything else, then re-test the tone-related rule. Move the tone you actually want into Custom Instructions so it is not at the mercy of a preset.
Step 5: Audit Project instructions for contradictions
If the failing chat is inside a Project, open the Project, click the three-dot menu, then Instructions. If they contradict your Custom Instructions, align them or move the chat out of the Project.
Step 6: Shorten and sharpen the instructions
Cut each field well under the 1,500-character cap. Replace vague rules with concrete, testable ones:
Bad: Be more concise and direct.
Good: Maximum 5 bullet points per reply. No introductory sentence.Save and test in a new chat.
Step 7: Pin your model choice
In the model picker, choose GPT-5.5 (Instant or Thinking) explicitly instead of leaving it on Auto. This reduces the kind of silent switch to GPT-5.4 mini that waters down instruction following. Note: once you hit the cap, ChatGPT may still fall back regardless of your pick.
Step 8: Reapply instructions inline for stubborn chats
In a long chat where instructions have drifted, re-anchor the model without restarting:
Reminder: from now on in this conversation, follow my account
Custom Instructions exactly. In particular: max 5 bullets, no
introductory sentence, 10th-grade reading level.Step 9: For team-wide consistency, move rules into a Project
If multiple chats need the same behavior, create a Project, put the rules into the Project’s Instructions, and run those chats inside it. Project instructions are obeyed more reliably than account-level ones for chats inside that Project.
How to confirm it is fixed
- A new chat outside any Project follows the rules from turn 1.
- A new chat inside a Project follows the merged Project plus account rules.
- After ~20 turns and one file upload, the rules still hold.
- The one rule that was being ignored (the Memory or Personality conflict) now sticks in a fresh chat.
- Switching from desktop to mobile and back keeps behavior consistent.
Long-term prevention
- Keep account-level Custom Instructions short, focused on tone and format only.
- Keep Memory clean: every contradicting fact you save is a future override. Audit
Manage memoriesmonthly. - Put workflow-specific rules into Projects, not at the account level.
- Use hard, testable rules (“never use the word ‘leverage’”) instead of fuzzy ones.
- Pin the model in the picker for important sessions; do not rely on Auto.
Common pitfalls
- Packing both fields to the 1,500-character cap and expecting every line followed equally.
- Mixing contradicting rules (“be concise” and “always give 3 detailed examples”).
- Forgetting that a saved Memory or a non-Default Personality preset can quietly outrank an instruction.
- Assuming Project and account instructions stack additively; they can conflict.
- Editing Custom Instructions on mobile and expecting instant sync on desktop.
FAQ
- What is the priority order? Roughly: inline prompt in the current turn > Project instructions > account Custom Instructions and saved Memory > Personality preset > model defaults. When Memory and Custom Instructions conflict, the more specific or more recently reinforced one tends to win, which is why a stale Memory can override a rule.
- Why does even a brand-new chat ignore one rule? Almost always a saved Memory that contradicts it, or a Personality preset overriding tone. Fresh chats clear conversation context but not Memory or presets.
- How long can Custom Instructions be? Each field caps at 1,500 characters (3,000 total) as of June 2026. Shorter is followed more reliably.
- Do Custom Instructions apply to API calls? No. The API uses the system prompt you set in code; the web Personalization settings do not carry over.
- Does Memory replace Custom Instructions? No. Memory holds facts about you across chats; Custom Instructions set rules. They coexist, and they can also conflict, which is the trap.
- Can I have different Custom Instructions per workspace? Account-level applies everywhere; for per-workspace rules use Projects.