ChatGPT Custom Instructions Not Working? Fix Them Fast

What Custom Instructions Can and Cannot Control

They are guidance for personalization, not an absolute command layer. ChatGPT still follows a hierarchy of instructions. System-level rules, product behavior, developer instructions, and safety requirements can take priority over your custom instructions.

This is why ChatGPT may appear to ignore an instruction even when the setting is enabled. It may be following a higher-priority requirement rather than malfunctioning.

Check Your ChatGPT Personalization and Context Settings

Check that:

• Custom instructions are turned on.
• Your latest edits were saved.
• You edited the correct account.
• The instruction appears in the field after reopening settings.

After making changes, start a new standard chat. An existing conversation may still contain earlier context that conflicts with the updated preference.

That distinction can produce confusing results. A saved memory may support, compete with, or appear to override a custom instruction. Review your memory controls and saved memories if ChatGPT repeatedly uses outdated information or preferences.

Also check whether you are using Temporary Chat. Temporary chats have different personalization and memory behavior, so they may not reflect your usual settings in the same way as a standard conversation.

Finally, consider the conversation itself. Earlier messages can influence the current answer. If you told ChatGPT to write in a formal style at the start of a long conversation, then later changed your custom instructions to request a casual style, the old chat may continue to reflect the earlier context.

Rewrite Instructions So ChatGPT Can Follow Them Reliably

Long, vague, or contradictory instructions are difficult to apply consistently. “Be more useful” is subjective. “Use five bullet points and include one example” is observable.

Use short rules that describe behavior you can check. For example:

• Replace “Give better answers” with “Lead with the direct answer.”
• Replace “Keep it readable” with “Use headings and paragraphs of no more than three sentences.”
• Replace “Always be detailed” with “Explain unfamiliar terms and include one practical example.”

Avoid putting conflicting rules in the same instruction set. “Be concise” and “Explain every possible edge case” may both be reasonable, but they need a priority rule such as: “Be concise by default; add edge cases only when they could change the recommendation.”

Role: Describe the perspective or working style you want.

Response format: Specify headings, bullets, tables, length, or ordering.

Must do: List two or three behaviors that matter most.

Avoid: Identify habits you do not want, such as repetition or unnecessary introductions.

Example: Show one short example of the desired output.

For example:

Role: Act as a practical writing editor.

Response format: Use headings and bullet points.

Must do: Lead with the main recommendation and explain unfamiliar terms.

Avoid: Do not repeat my question or add a generic conclusion.

Example: For a product comparison, start with the best choice, then give three reasons.

Prioritize the rules that matter most. A compact instruction set is easier to test and less likely to contain accidental conflicts. If the setting has a character limit, remove repetition, combine related rules, and keep only preferences you genuinely use.

A Step-by-Step Test for Custom Instructions That Are Not Working

Use this diagnostic process before rewriting everything at once.

You can also ask ChatGPT to summarize the preferences it is expected to follow. Treat that summary as a diagnostic clue, not proof that every rule will be applied perfectly.

Fix Client and Account Problems Before Escalating

A personalization issue may be caused by the client rather than the instruction. Try the following:

• Refresh the page.
• Sign out and sign back in.
• Update the ChatGPT mobile app.
• Test another supported browser or device.
• Confirm you are using the intended account.
• Check the service status page for an active incident.

Repeat the fresh-chat test after each meaningful change. This prevents you from confusing several fixes performed at once.

When to Contact OpenAI Support

If the same simple instruction fails in a new standard chat across supported clients, contact OpenAI Support with useful diagnostic details. Include:

• The exact custom instruction.
• The exact test prompt.
• The expected output and actual output.
• The model and mode used.
• The client, operating system, and app or browser version.
• Whether the chat was temporary or standard.
• The date and time, including your time zone.
• The steps you already tried.

Start with the fresh-chat diagnostic, then rewrite your instructions as short, testable rules. In many cases, the cause is an old conversation, memory behavior, a temporary chat, or an unclear rule - not a broken setting. If the failure persists across supported clients, provide the complete test details when you contact Support.

Step-by-Step Guide

1. 1

   Test one instruction in a new standard chat

   Enable custom instructions, open a new non-temporary chat, and test one harmless preference such as answering in three bullet points.

2. 2

   Compare fresh and existing conversations

   Run the same prompt in a new chat and the affected conversation to identify whether earlier messages are overriding or influencing the preference.

3. 3

   Check personalization and context settings

   Confirm that custom instructions are enabled and saved, then review memory, Temporary Chat, account, model, and mode settings.

4. 4

   Simplify and rebuild the instructions

   Remove all but one rule, rewrite it as a clear observable behavior, and add the remaining rules back gradually while removing contradictions.

5. 5

   Check the client before escalating

   Refresh the page, sign in again, update the app, try another supported browser or device, and check the service status page.

6. 6

   Contact OpenAI Support with evidence

   Provide the exact instruction, test prompt, expected and actual results, model, mode, client details, timestamp, and troubleshooting steps already completed.

Key Statistics

• The article defines a four-part instruction template: Role, Response format, Must do, and Avoid, with an optional example.Partnerin AI's recommended structure turns broad preferences into testable rules that are easier to apply and troubleshoot.
• The diagnostic workflow compares two conversation conditions: a new standard chat and the existing affected chat.Partnerin AI recommends this controlled comparison to identify whether earlier conversation context is influencing the response.
• The support checklist contains eight key diagnostic details, including the instruction, test prompt, expected result, actual result, model, client, timestamp, and prior troubleshooting.Partnerin AI's escalation guidance is designed to give OpenAI Support enough context to reproduce a persistent personalization issue.

Frequently Asked Questions