Why ChatGPT Ignores Your Custom Instructions (And How to Fix It)

You set up custom instructions. ChatGPT follows them... for 2 messages. Then ignores them completely.

Related troubleshooting guides: [stopping verbose responses](/blog/stop-chatgpt-verbose), [getting specific answers](/blog/chatgpt-generic-answers), [avoiding fake citations](/blog/chatgpt-fake-citations), and [improving your prompts overall](/blog/write-better-prompts).

The Real Problem

Custom instructions are supposed to persist across conversations. But many users report that ChatGPT stops following them mid-conversation. Here's why it happens and exactly how to fix it.

Why ChatGPT Stops Following Custom Instructions

• Context window limit: Long conversations push custom instructions out of context
• Token overflow: Your instructions + conversation = exceeds token limit
• Model context switching: Using different tabs/windows resets context
• API calls: Directly calling API may not include custom instructions
• Instruction conflicts: New instructions override custom ones
• Session changes: Switching browsers or devices may not sync instructions

Root Cause #1: Context Window Limits

ChatGPT includes your custom instructions at the beginning of each conversation. As your conversation grows, the model has less space for instructions. Eventually, it deprioritizes them.

Example: If you have 2000 tokens of instructions + 6000 tokens of conversation history, the model may choose to keep recent messages instead of your full instructions.

Root Cause #2: Instruction Conflicts

When you give a new instruction in the chat, it conflicts with your custom instructions. The model prioritizes recent instructions over background ones.

Your custom instruction says: 'Always use formal tone.' You then ask: 'Write this casually.' ChatGPT chooses the immediate instruction over the background one.

The Fixes That Work

Fix #1: Keep Instructions Short and Specific

Long instructions dilute themselves. Prioritize essential behavior over nice-to-haves.

Good: 'Always end responses with an action step.' Bad: 'You are a helpful assistant who speaks with warmth, enthusiasm, and always considers multiple perspectives while being mindful of nuance and edge cases.'

Keep custom instructions under 500 tokens. If you have more, you're probably mixing strategy with tactics.

Fix #2: Use Role-Based Instructions Instead of Behavior

Instead of listing behaviors, define a role. Roles are stickier than instructions.

❌ WEAK:
Always be concise. Always provide examples. Always end with a summary.

✅ STRONG:
You are a technical writer for developers. You explain concepts clearly with code examples.

Role-based instructions anchor the model's behavior without listing every expected action.

Fix #3: Reinforce Instructions in Long Conversations

For conversations over 20 messages, remind ChatGPT of key instructions every 10-15 messages.

User message: 'Remember, I need all code examples in Python with comments.'

This resets the context window and makes your instructions stick.

Fix #4: Use the System Prompt Pattern

In each message, include a mini system prompt that reinforces your core requirement.

Your message:
'[System: Respond in a friendly, conversational tone. Avoid jargon.] Now, explain how neural networks work.'

Fix #5: Start Fresh Conversations for New Topics

Custom instructions work best at conversation start. If a conversation drifts far from the original topic, start a new conversation.

This resets the context, reloads your custom instructions at full strength, and gives ChatGPT a fresh frame.

Template: The Sticky Custom Instruction

Use this pattern for custom instructions that actually stick:

1. [Define your role/persona in one sentence]

2. [State your core requirement - the ONE thing that matters most]

3. [Give one specific format/style example]

Example:
'You are a technical writer. Core requirement: Always provide working code examples before explanations. Style: Conversational tone, avoid corporate jargon. Example output format: [Code] then [Explanation].'

Quick Checklist

• ☐ Custom instructions are under 500 tokens
• ☐ Instructions define a role, not just behaviors
• ☐ For long conversations, reinforce key points every 10-15 messages
• ☐ Start fresh conversations when topic changes significantly
• ☐ Test instructions in a new conversation to verify they work
• ☐ If still failing: switch to API with explicit system prompt