The Ultimate Guide to ChatGPT Custom Instructions 2026: Make AI Always Remember Who You Are

Most people reintroduce themselves in every new chat—"I'm a backend engineer," "Keep answers concise," "Use Chinese." Custom Instructions, configured once and automatically applied to all new conversations, is the most cost-effective and underrated feature in ChatGPT. This article provides templates for the two input boxes, ready-to-use configs by profession, how to combine with Memory/Projects, and iteration methods.

1. What to Write in Each Box

Access: Avatar → Settings → Personalization → Custom Instructions.

First box "About You" = static background (who you are, what you know, what you want):

text
I'm a backend engineer with 5 years of experience, primarily using Python and Go, familiar with AWS/K8s, and have basic knowledge of ML.
Currently working on the API layer of an AI SaaS product.
Preferences: explain principles before giving code; when making technical choices, I want to see multi-option comparisons with a clear recommendation.

Second box "How You Want to Be Answered" = behavior rules (how to respond):

text
Give the conclusion first, then the reasoning. No polite opening lines or "hope this helps" endings.
Code defaults to Python 3.12 with type annotations; provide complete runnable examples, not snippets.
If unsure, say "I'm not sure" explicitly; never fabricate APIs/parameters/data.
Point out obvious flaws in my approach directly, don't pander.
When multiple approaches exist, provide a comparison table + your recommendation.

Key writing principles: Be specific rather than vague ("5 years Python backend" >> "programmer"); write rules as actionable imperative sentences ("Don't X", "Default to Y")—the model follows negative lists most reliably. This aligns with general prompt engineering principles: explicit rules drastically reduce output variance.

2. Ready-to-Use Templates by Profession

Product Manager:

text
[About You] B2B SaaS product manager, responsible for enterprise tool growth, with data analysis background, familiar with SQL/A-B testing.
[How to Answer] Output should be structured (bullet points/tables); for feature design, first ask about target users and scenarios before answering;
when providing competitor comparisons, note that info may be outdated and needs verification.

Content Creator:

text
[About You] Self-media writer, mainly writing tech reviews, audience is non-tech enthusiasts, platforms are WeChat Official Accounts and Xiaohongshu.
[How to Answer] Writing style should be conversational but not cheesy; provide 5 headline alternatives; any jargon must be explained in plain language;
by default, output two versions: one with emojis for Xiaohongshu and one without for WeChat Official Accounts.

Student/Researcher:

text
[About You] Graduate student, research direction NLP, read many English papers but writing is average.
[How to Answer] Explain concepts in three parts: "intuitive explanation → formal definition → an example";
when helping me revise academic English writing, give reasons for changes; when citing papers, provide source and remind me to verify.

3. How to Divide Work with Memory and Projects

Each mechanism handles its own scope—don't mix them:

MechanismWhat It ManagesCharacteristics

Custom InstructionsLong-term stable identity and response styleGlobal, manually maintained, most controllable MemoryScattered facts naturally accumulated in conversationsAuto-recorded, occasionally wrong—periodically clean incorrect memories in settings ProjectsProject-specific context and filesEffective within a project, suitable for local rules like "this project uses Vue, not React"

A common mistake is stuffing project details into global instructions—every response gets polluted after switching projects. Global for constants, project for variables.

4. Iteration Method: Treat It Like Prompt Engineering

After configuration, test with 5 typical questions and observe which rules didn't take effect.

Make ineffective rules more specific ("concise" → "default to under 300 words, say 'elaborate' to lengthen").

Every time you encounter an unsatisfactory response pattern, distill it into a new rule (prioritize high-frequency pain points within the limit).

Review every 1-2 months—your tech stack and needs change.

The experience gap typically becomes noticeable after about 3 rounds of iteration: from "generic assistant" to "a colleague who knows you."

FAQ

Q: Does it affect existing conversations? Only new conversations; after editing instructions, start a new chat to test.

Q: Do Claude/Gemini have equivalents? Yes, similar mechanisms exist (Claude's Profile/Projects, Gemini's Saved Info), and the writing principles are universal—the templates in this article can be directly ported.

Q: How does it relate to system prompts? Custom Instructions are essentially a user-controllable system prompt overlay, so all prompt engineering techniques (specific, actionable, negative lists) apply.

---

*Last updated: June 2026. Feature access and limitations are subject to ChatGPT's official documentation.*