ChatGPT Custom Instructions Not Working? Practical Fixes and Diagnostic Tests
TL;DR Check the account, saved custom-instruction fields, and new-chat behavior before rewriting your instructions. Use a clean browser session to separate ChatGPT settings problems from stale cookies, extensions, or app…
- Check the account, saved custom-instruction fields, and new-chat behavior before rewriting your instructions.
- Use a clean browser session to separate ChatGPT settings problems from stale cookies, extensions, or app state.
- Run the BUGI-TEST marker check to see whether custom instructions are being ignored everywhere or only in one surface.
Overview
When ChatGPT custom instructions seem broken, the failure is often smaller than it looks. The saved text may be attached to a different account, the current conversation may be carrying older context, or the browser session may not have refreshed correctly. Before changing a long instruction set, verify the surface you are using: same account, same workspace, new chat, and saved fields still present.
OpenAI documents the feature as a user-configurable behavior setting in OpenAI’s custom instructions help article. That does not mean every reply will behave like a hard rule. Custom instructions sit alongside the current prompt, model behavior, safety rules, memory, GPT configuration, and workspace controls.
- Open ChatGPT settings and confirm the custom instruction fields are present, enabled, and saved.
- Start a new chat after saving. Do not use an old conversation as the main test.
- Try an incognito or private window with extensions disabled.
- Compare the same account on another browser, device, or app surface.
- Temporarily reduce the instruction text to one obvious rule, then add the rest back after it works.
Symptoms
The clearest sign is a reply that ignores a stable preference you already saved: language, tone, coding style, formatting, role context, or a rule to avoid a specific behavior. Sometimes the settings panel still shows the instructions, but a new chat answers as if the fields are blank.
The pattern matters. If the issue appears only in the mobile app, treat it as an app-session problem first. If it appears only in a custom GPT, inspect that GPT’s own configuration. If it happens only in one browser profile, look for extensions, blocked cookies, or stale site data. One bad answer is not enough evidence for a service outage or account-wide failure.
Why This Happens
Custom instructions are persistent preferences, not a guaranteed command layer. ChatGPT still has to resolve the active user prompt, system behavior, tool-specific rules, safety constraints, memory behavior, and the product surface being used. A saved instruction such as “answer in concise technical English” can be weakened by a later prompt asking for a long beginner walkthrough.
Local state can also mimic a broken setting. Browser cache, blocked cookies, privacy extensions, account switching, stale mobile sessions, and managed workspace controls can all make saved instructions appear unreliable. The safest path is to isolate one variable at a time: account first, clean session second, instruction text third.
Check status, account, and basic access
Start with checks that do not change your instruction text. Confirm you are signed into the same ChatGPT account where the instructions were saved. Open the custom instructions panel, verify both fields still contain the expected text, then save again. Create a new chat and test there.
If ChatGPT itself is unavailable or degraded, use OpenAI’s official status page rather than guessing from social posts. If you use a work, school, Team, Enterprise, or other managed account, check whether workspace settings affect personalization. Do not assume a managed workspace behaves exactly like a personal account.
Verify the active account
Check the email or workspace shown in ChatGPT before editing custom instructions. Account switching is a common cause of “missing” settings.
Try a clean browser or app session
A clean session quickly separates a ChatGPT settings issue from a local browser issue. Open a private or incognito window, sign in, and create a new chat with a simple verification prompt such as “Reply in the exact format required by my custom instructions.”
If the clean session works, return to the main browser profile and disable extensions that modify pages, block scripts, rewrite requests, or manage cookies. On mobile, force-close the app, reopen it, and test again. If the problem appears only in one app surface, record that exact split before escalating.
Test without extensions
Use a private window or another browser profile. Ad blockers, privacy tools, and script managers can interfere with saved settings or session refresh.
Confirm the instructions are clear and not self-conflicting
Custom instructions can be saved correctly and still produce weak results if the text is vague, overloaded, or internally inconsistent. Put the highest-priority rule first. Use direct language: “Always answer in English,” “Use TypeScript unless I request another language,” or “Prefer concise bullet lists for troubleshooting.”
Avoid mixing broad identity prompts with operational rules in the same paragraph. A practical structure is role context, output format, coding preferences, and constraints. Keep negative rules specific. “Do not be verbose” is less useful than “Keep default answers under 150 words unless I ask for a deep dive.”
Reduce the instruction set
Temporarily keep only one formatting rule and one domain preference. If those work, add the rest back in small groups.
Check memory, temporary chats, GPTs, and project scope
Custom instructions are only one layer of ChatGPT personalization. Memory, temporary chat behavior, custom GPT configuration, project-level context, shared-chat context, and developer-facing instructions can change the final output. OpenAI describes memory separately in OpenAI’s Memory FAQ, which matters because users often confuse memory behavior with custom instruction behavior.
If you are using a custom GPT, check that GPT’s own instructions and configuration. OpenAI’s official guide to GPTs is available in OpenAI’s GPT creation help article. If you are using a temporary or restricted mode, test again in a normal new chat before deciding that custom instructions are broken everywhere. For coding workflows, prompt-level repository context can override broad style preferences.
Separate personalization layers
Test one normal new chat, one temporary chat if available, and one custom GPT if relevant. Compare which surface ignores the instruction.
Run platform-specific checks
For Chrome, Edge, and Brave, test with all extensions disabled and confirm cookies are allowed for ChatGPT. For Safari, disable content blockers for the site and test outside Private Browsing. On iPhone or Android, compare the app with the web version in a mobile browser using the same account.
On desktop apps, sign out and back in before reinstalling. Reinstallation should be a later step, not the first fix. If your custom instructions work on web but not mobile, preserve that exact platform split when contacting support.
Compare web, mobile, and desktop
Use the same account and the same short test prompt across surfaces. The result tells you whether the issue is local or account-wide.
Use a minimal reproducible prompt
Do not troubleshoot with a complex work prompt. Create a short custom instruction such as “Always start your answer with BUGI-TEST.” Save it, open a new chat, and ask: “What is 2 plus 2?”
If the response starts with the marker, the feature is applying and your original instruction set likely has a conflict or priority problem. If the marker is ignored across browsers and devices, the problem is more likely account, product-surface, or service related. Remove the marker after testing so it does not pollute future conversations.
Temporary custom instruction test:
Always start your answer with BUGI-TEST.
Test prompt:
What is 2 plus 2?
Expected signal:
The answer starts with BUGI-TEST.
When Custom Instructions Are the Wrong Tool
Use custom instructions for stable preferences: language, tone, coding style, naming conventions, response format, and recurring context. Avoid using them for one-off task requirements, secrets, live project state, or instructions that change daily.
For coding work, put repository-specific rules in the current prompt, project context, or tool-native rules where available. Use ChatGPT custom instructions for default behavior, not compliance-critical behavior. If a rule must be followed every time for production work, repeat it in the active prompt or encode it in your development workflow.
If It Still Fails
If the issue persists after clean-session, account, and minimal-prompt checks, collect a compact support packet. Include the account type, platform, browser or app surface, whether the same instructions work on another device, and a sanitized screenshot of the saved instruction panel.
Capture the minimal test prompt and response. Do not include secrets. Then use OpenAI’s help or support route from the ChatGPT interface. Keep a local copy of your custom instructions before making further changes, because repeated edits can hide the original failure pattern.
FAQ
Why are my ChatGPT custom instructions not working?
They may not be saved in the active account, the browser session may be stale, the current chat may contain conflicting instructions, or the surface you are using may handle personalization differently. Start with a new chat, a clean browser session, and a minimal test instruction.
Do custom instructions apply to old chats?
Do not rely on old conversations as the main test. Start a new chat after saving changes. Existing context in an older conversation can compete with or dilute newly saved preferences.
Can ChatGPT ignore custom instructions?
Yes. Custom instructions are preferences, not an absolute override. Higher-priority system rules, safety behavior, tool-specific instructions, custom GPT configuration, and the active user prompt can affect the final answer.
Should I clear cache to fix custom instructions?
Try a private window or another browser first. If that works, then clear site data or disable extensions in the main browser profile. Clearing all browser data is broader than necessary for most users.
Why do custom instructions work on web but not mobile?
That points to a platform-specific session or app issue. Sign out and back in on the mobile app, force-close it, and compare the same account in a mobile browser before reinstalling.
What is the best test for custom instructions?
Use one obvious temporary rule, such as “Always start your answer with BUGI-TEST,” then ask a simple question in a new chat. This isolates the feature from complex prompt behavior.
Should I put project-specific coding rules in custom instructions?
Only if they are stable across most chats. For repository-specific conventions, use the active prompt, project context, or tool-native configuration instead. Custom instructions are best for default preferences, not changing project state.