What Custom Instructions Actually Do
Custom Instructions inject a hidden system prompt into every new conversation you start. You write it once; ChatGPT reads it before every reply. This eliminates repetitive setup like:
- "I am a marketing manager, so keep the tone professional"
- "Give me bullet points, not long paragraphs"
- "I am based in Australia, use Australian English"
- "Skip the 'Great question!' opener"
Without Custom Instructions, each new chat starts blank. With them, ChatGPT always knows who you are and how you want it to behave.
How to Set Them Up
- Click your avatar or name in the bottom-left of the ChatGPT sidebar
- Click Settings
- Go to Personalization
- Click Custom Instructions
- Fill in both text boxes (details below)
- Click Save
Changes take effect on your next new conversation. They do not retroactively change ongoing chats.
What to Write: Box 1 — "What should ChatGPT know about you?"
This box is for context. Write facts that would change how ChatGPT answers you.
Template structure:
Role / profession: [what you do]
Expertise level: [beginner / intermediate / expert in X]
Main use cases: [what you typically use ChatGPT for]
Location / language: [if relevant]
Recurring project context: [if you work on something specific regularly]
Example for a developer:
I am a backend software engineer with 8 years of experience, primarily in Python and Go.
I mainly use ChatGPT for code review, architecture decisions, and debugging.
I am building a SaaS product with a FastAPI backend and PostgreSQL database.
Assume I understand programming concepts — no need to explain basics.
Example for a content creator:
I run a YouTube channel and newsletter about personal finance for people in their 30s.
I use ChatGPT for scripting videos, writing email copy, and brainstorming content ideas.
My audience is educated but not finance professionals — clear, jargon-free language is essential.
I am based in Canada.
What to Write: Box 2 — "How should ChatGPT respond?"
This box controls style, tone, format, and behavior.
High-impact instructions to include:
- Length: "Keep answers concise. Use the shortest response that fully answers the question."
- Format: "Use bullet points for lists. Use numbered steps for processes. Avoid walls of text."
- Tone: "Use a direct, professional tone. Avoid overly casual language."
- Openers to remove: "Do not start with 'Certainly!', 'Great question!', or 'Absolutely!'"
- Closers to remove: "Do not add motivational sign-offs or suggest I ask follow-up questions."
- Caveats: "Skip safety disclaimers unless there is genuine risk involved."
- Language: "Use British English spelling and punctuation."
- Code: "When writing code, include comments only for non-obvious logic. Use type hints in Python."
Example (developer):
Be concise and technical. Assume expert-level familiarity with software engineering.
Use code blocks for all code. Include type hints in Python. Skip basic explanations.
Do not add caveats about consulting professionals.
Start responses directly — no 'Of course!' or 'Sure!' openers.
Example (content creator):
Write in a warm, conversational tone that feels human, not corporate.
Keep paragraphs short (2–3 sentences max). Use bullet points for lists.
When I ask for ideas, give at least 8 options.
Do not use jargon. No preamble — get to the point immediately.
Specific Instructions That Make a Big Difference
These are high-value additions most people miss:
Stop unnecessary hedging:
"Do not add phrases like 'it is important to note that' or 'it is worth mentioning.' State information directly."
Remove boilerplate advice:
"When I ask a coding question, do not suggest I hire a professional or check the documentation unless I specifically ask for resources."
Control output length by task:
"For quick factual questions, give short answers. For creative or writing tasks, give complete, polished output without asking me to confirm length first."
Force consistent formatting:
"Always use markdown formatting. Use headers for sections longer than 200 words."
Common Mistakes to Avoid
- Writing too much — Keep each box under 1500 characters. Long instructions get partially ignored. Prioritize the most important 3–5 rules.
- Being vague — "Be helpful" or "be smart" adds nothing. Be specific: "answer in under 100 words for factual questions."
- Contradicting yourself — If you say "be concise" in one line and "give thorough explanations" in the next, ChatGPT will struggle to reconcile them. Pick one as the default.
- Forgetting to save — Click Save before closing the settings panel. There is no auto-save.
- Expecting mid-conversation updates to apply — Instructions only apply at conversation start. Update them, then open a new chat.
- Not testing — After saving, open a new chat and ask a question you usually repeat context for. Check whether ChatGPT's response reflects your instructions without prompting.
Example: Before and After Custom Instructions
Without custom instructions:
You: "Explain REST APIs." ChatGPT: [3 paragraphs of basics starting with "A REST API, or Representational State Transfer..."]
With custom instructions (backend dev, skip basics):
You: "Explain REST APIs." ChatGPT: [2 sentences covering statelessness, HTTP verbs, and resource-based URLs — then asks what aspect you want to go deeper on]
The same prompt, completely different output quality.