Imagine briefing a brilliant assistant once — who you are, how you work, what you want — and never repeating yourself. That’s custom instructions: the permanent context layer that shapes every ChatGPT response. How to set custom instructions in ChatGPT? It’s a five-minute setup with compounding returns. Most users never touch it; the ones who do never go back.
How to set custom instructions in ChatGPT
What custom instructions actually do
Two boxes, per OpenAI’s custom instructions guide: “What would you like ChatGPT to know about you?” (your context — role, background, projects, preferences) and “How would you like ChatGPT to respond?” (your style rules — tone, format, depth, things to avoid). These prepend to every conversation automatically. The effect: no more “I’m a high school teacher, explain simply” in every prompt — it’s just known. It’s the highest-leverage five minutes in ChatGPT power usage.
How to set them (the clicks)
- Click your profile → “Customize ChatGPT” (wording varies slightly by platform/version).
- Fill in the two boxes (guidance below — this is where the value lives).
- Save. They apply to new conversations going forward.
- Toggle per-chat if needed — some conversations want a clean slate (the option exists; use it for role-play or when the instructions would interfere).
What to put in box 1: who you are
Include: your role and industry, key projects or contexts, your expertise level (so it calibrates — expert vs. beginner explanations), relevant constraints (tools you use, standards you follow), and location/timezone if relevant. Example: “I’m a product manager at a B2B SaaS company (50 employees). I work on onboarding and activation. Technical enough to read code, not to write it. Based in Berlin.” Don’t include: secrets, credentials, highly sensitive personal data — this is stored server-side and prepended to everything. Treat it as semi-public.
What to put in box 2: how to respond
This is the craft. Effective instructions are specific and behavioral:
- Format defaults: “Use headers and bullet points for anything over 200 words. Lead with the answer, then the reasoning.”
- Depth calibration: “Assume I know the basics; skip introductory explanations unless I ask.”
- Tone: “Direct and candid. Don’t sugarcoat; tell me when my idea is bad.”
- Behavioral rules: “When I ask for options, give 3 with trade-offs and a recommendation. Always include the counterargument.”
- Prohibitions: “Never use em-dashes. No inspirational closing paragraphs. Don’t ask if I want more help at the end.”
- Domain specifics: “For code: explain the why, not just the what. For writing: match my casual-but-precise voice.”
The pattern: rules you’d give a human assistant after working together for a month. Write them as that briefing.
Examples for common profiles
The developer: Box 1: “Senior backend engineer, Python/Go, distributed systems.” Box 2: “Code-first answers. Explain trade-offs, not syntax basics. Flag security implications proactively. Concise.”
The writer: Box 1: “Freelance copywriter, B2B tech clients.” Box 2: “Draft in my voice: direct, wry, no corporate speak. Ban overused explorer verbs, ‘landscape’ metaphors, and summary-announcement closers. Always give 3 headline options.”
The student: Box 1: “Second-year biology undergrad.” Box 2: “Socratic style — guide, don’t just answer. Connect new concepts to what we’ve covered. Quiz me after explaining.”
The executive: Box 1: “COO of a 200-person logistics company.” Box 2: “Bottom line first, always. Quantify when possible. Flag risks explicitly. No fluff — I read fast.”
Custom instructions vs. memory vs. projects
Three overlapping features, different jobs: custom instructions = permanent briefing (who you are, how to respond — rarely changes). Memory = accumulated facts (things learned about you over time — evolves; see ChatGPT memory). Projects = scoped context (this project’s files and instructions — see make ChatGPT remember context). Use all three: instructions for the permanent, memory for the accumulated, projects for the scoped. They compose beautifully.
Maintenance
Instructions for teams (shared conventions)
If your team uses ChatGPT, consider shared custom-instruction conventions: a team baseline everyone adopts (company voice, “cite sources for factual claims,” “flag uncertainty explicitly”) plus personal additions. The benefits compound — consistent output quality across the team, onboarding that’s instant (“here are our AI settings”), and a shared language for improving prompts. Keep the team conventions in a shared doc, review quarterly, and let individuals customize on top. One caution: don’t over-standardize. The marketing team’s instructions differ legitimately from engineering’s — shared baseline, not straitjacket. And for truly shared contexts (not just instructions), team projects/workspaces are the right layer — instructions are personal defaults, projects are shared work areas. Use each for its job.
Review quarterly — roles change, projects end, preferences evolve. Stale instructions are worse than none (the AI confidently personalizing for a job you left). And when answers feel off, check the instructions first — a well-meaning rule from six months ago is often the culprit. Five minutes to set up, two minutes quarterly to maintain, better answers forever. OpenAI keeps refining how these interact with memory and projects; the get better answers from ChatGPT guide covers the per-prompt side of the same skill. Set the briefing once; benefit every day.
1 Comment