ChatGPT Custom Instructions Guide: System-Level Prompting

ChatGPT Custom Instructions are the account-level settings that tell the model who you are and how every reply should look. You fill them once under Settings → Personalization (or Customize ChatGPT on mobile), turn Enable customization on, and those rules ride along with new chats on Free, Plus, Team, and Enterprise.

This guide treats Custom Instructions as a product feature: where the boxes live, what belongs in each field, how they stack with Memory, Custom GPTs, and Projects, and how to test them on GPT-5.5 Instant (the usual fast ChatGPT default as of mid-2026). You leave with field recipes, a setup checklist, and failure modes that show up when the boxes fill with personality essays.

What ChatGPT Custom Instructions are

Custom Instructions are persistent preferences attached to your ChatGPT account. They are not a separate product, a Custom GPT, or an API system message you paste per call. OpenAI applies them as background context so you stop retyping the same role, audience, and format rules at the start of every thread.

On web and desktop, open your profile menu, choose Settings, then Personalization, then Custom Instructions. On iOS and Android, open Settings and choose Customize ChatGPT. Confirm Enable customization is on. As of mid-2026 the UI still centers on two text fields: what ChatGPT should know about you, and how it should respond. Style or tone presets (Default, Professional, Friendly, Candid, and similar) may sit nearby; they stack with your written text rather than replace it.

Each field is short. Plan for about 1,500 characters per box. That limit is a feature: it forces you to keep stable facts and a few hard rules instead of a novel. If your draft overflows, cut temporary project names, move client-specific voice into a Project or Custom GPT, and keep the global boxes for preferences that stay true for months.

Who this helps: people who live in ChatGPT daily, freelancers who need a consistent reply shape, developers who want code-first answers, and anyone tired of fighting "Great question!" openers. Who should keep the boxes light: users who jump across unrelated jobs in the same account without Temporary Chat, and teams that need per-client rules (those belong in Projects or Custom GPTs, not the global account layer).

How Custom Instructions work (and what they do not do)

When customization is enabled, ChatGPT includes your saved text as part of the context for chats that respect personalization. OpenAI's help docs state updates apply across chats, including ones already open, once you save. Still treat a fresh thread as the clean test environment: old messages in a long conversation can pull the model toward earlier habits even when your settings changed five minutes ago.

Think of three layers. Layer one is Custom Instructions: stable identity and default response rules. Layer two is Memory, when it is on for your plan: facts ChatGPT learned across conversations and may recall later. Layer three is the current user message: the task for this turn. Custom Instructions set the baseline. Memory fills in recent context. The user message assigns the job. Mixing those jobs is what makes replies feel random.

Custom Instructions do not create a private vault. Anything you paste can sit in the same trust boundary as other chat text. Skip passwords, API keys, payroll numbers, medical details, and client secrets. If you care about training opt-out on consumer plans, set that under Data Controls; that choice is separate from whether customization is enabled. Temporary Chat (or equivalent ephemeral mode) is the escape hatch when you want a reply that ignores both Custom Instructions and Memory.

Field 1: what ChatGPT should know about you

Field 1 is context about you, not a biography and not a list of how the model should talk. Put role, domain, tools, audience, and constraints that stay true across most chats. Leave out one-week project names unless you live inside that project all day.

Strong Field 1 ingredients:

• Role and seniority in one line ("senior product marketer at a B2B SaaS company")
• Domains you work in most days ("lifecycle email, landing pages, experiment design")
• Audience defaults ("I write for technical buyers, not consumers")
• Tooling and stack when it changes answer shape ("TypeScript, Next.js, Postgres")
• Hard boundaries that are about you ("I am not a lawyer; flag when a question needs counsel")

Weak Field 1: hobbies, life story, motivational slogans, and "I love AI." Those burn character budget and almost never change output quality. Also weak: pasting a full resume. The model needs working context, not a CV parser demo.

Field 2: how ChatGPT should respond

Field 2 is your standing response contract. This is where format, tone, length, and refusal habits live. Treat it as the account-level version of a system-style control plane, still limited by the character cap. If you need a six-section production agent prompt for a Custom GPT or API worker, that longer structure belongs elsewhere; our separate system prompt guide covers that pattern. Custom Instructions Field 2 should stay short and enforceable.

Strong Field 2 ingredients:

• Opening rule ("Start with the answer. No praise of the question.")
• Default length ("Under 200 words unless I ask for depth")
• Format defaults ("Bullets for options; short prose for explanations; code in fenced blocks")
• Uncertainty rule ("Label guesses as guesses; ask one clarifying question when the brief lacks audience")
• Scope redirect ("If I ask for medical diagnosis, decline and suggest a clinician")

Keep Field 2 free of task-specific workflows. "Always produce a SWOT for any company I name" will fight you the day you ask for a joke or a regex. Put task workflows in the chat, a Project, or a Custom GPT.

Step-by-step: set up and pressure-test your instructions

Setup is short. Quality comes from the write-test-cut loop after you save. Most people fill both boxes in five minutes, run one happy-path question, and declare victory. That misses contradictions, character overflow, and rules that only work on the first reply. Budget another ten minutes for a small stress suite before you trust the settings for work.

Write offline first if you edit at a slow pace. Draft both fields in a notes app, count characters, then paste. If the blank Field 2 box stalls you, sketch the outcome you want in plain language and tighten it with PromptMake /text as a draft pass, then cut anything that sounds like filler before you save inside ChatGPT. The free tier is enough for a couple of drafts; you still own the final text that lands in Personalization.

Steps 1-3: open settings, enable, draft Field 1

1. Open ChatGPT on web or desktop. Click your profile, then Settings → Personalization → Custom Instructions. On mobile, use Settings → Customize ChatGPT.
2. Toggle Enable customization on. If the toggle is off, saved text does nothing.
3. Draft Field 1 to under ~1,500 characters. Use the role, domain, audience, stack, and hard boundary pattern above. Read it aloud once; if a sentence would not help ChatGPT choose vocabulary or depth, delete it.

Save when the UI offers a save control, or confirm autosave behavior in your client. Do not assume an unsaved draft is live.

Steps 4-6: draft Field 2, save, run a five-prompt test

1. Draft Field 2 with opening, length, format, uncertainty, and scope rules. Kill contradictions ("be thorough" next to "keep every answer under 50 words" without a scope line).
2. Save. Start a new chat for testing.
3. Run five prompts: (a) a short factual question in your domain, (b) an email draft, (c) a comparison with a requested table, (d) a vague request with missing audience, (e) an off-scope ask that should hit your redirect. Score each reply against Field 2. Edit the failing rule, save again, and retest only the failed cases.

Pass criteria: the opener rule holds, format matches without nagging, the vague case triggers a clarifier or an explicit assumption label, and the off-scope case redirects. If style presets are set to Friendly while Field 2 bans chattiness, change the preset or rewrite Field 2 so they agree.

Copy-paste recipes for both fields

Use these as starting points. Swap the bracketed details for your real role, audience, and stack. Keep each field under the character limit after edits. Prefer one job identity over three stacked personas that fight each other in the same box.

Read a recipe as a template, not a finished contract. After you paste, cut any line that does not change vocabulary, length, or format. Run the five-prompt stress test from the previous section before you trust the text at work. If a recipe field pushes past the character cap, delete biography and keep rules that you can score in a reply.

The three recipes below cover common PromptMake readers: product marketer, software engineer, and researcher. Pick the closest match, then rewrite Field 1 so it sounds like you and Field 2 so every rule is enforceable. When two recipes both fit, merge only the constraints you will actually enforce, not every bullet from both.

Recipe A: product marketer (Field 1 + Field 2)

Field 1: "I am a product marketer at a B2B SaaS company selling to engineering and IT buyers. I draft landing pages, lifecycle email, and launch briefs. Default audience is technical decision-makers. Stack context when relevant: web apps, APIs, usage-based pricing. Flag legal or compliance claims I should verify with counsel."

Field 2: "Start with the answer or the draft. No greetings and no praise of the question. Default under 200 words unless I ask for long form. Use short paragraphs or bullets. For copy drafts, return the copy first, then up to five change notes. If audience or offer is missing, ask one clarifying question before drafting. Do not invent customer quotes or metrics."

Recipe B: software engineer (Field 1 + Field 2)

Field 1: "I am a full-stack engineer working in TypeScript, Node, and Postgres. I prefer production-ready examples over pseudocode. Assume I can read docs; skip beginner tutorials unless I ask. Call out security and data-loss risks when they appear."

Field 2: "Lead with the fix or the code. Use fenced code blocks with language tags. Keep prose tight. If multiple approaches exist, give your pick first and one alternative in a short bullet. Ask one question when requirements are ambiguous. Refuse requests that need production credentials or live secret values."

Recipe C: researcher / analyst (Field 1 + Field 2)

Field 1: "I research markets and products for strategy memos. I need sources separated from inference. I write for executives who want bottom-line first."

Field 2: "Open with a two-sentence bottom line. Then give 3-5 bullets. Mark inferences as [INFERENCE]. When you lack sources, say so and list what you would verify. Skip motivational closers. Prefer tables for comparisons."

Custom Instructions vs Memory, Custom GPTs, and Projects

People treat these four features as synonyms. They are not. Pick the layer that matches how often the rule changes and how many people share it.

Custom Instructions: global defaults for your account. Best for identity, tone, and format that should apply almost everywhere. Change them when your job or standing preferences change, not when a single client brief changes.

Memory: learned or saved facts that can evolve across chats when the feature is available on your plan. Best for ongoing project names, preferences you stated in conversation, and durable notes you want recalled. Bad place for format contracts you need to enforce every time; put those in Field 2 so you can edit them in one screen.

Custom GPTs: specialized assistants with their own Instructions box, optional knowledge files, and tools. Best for repeatable workflows ("meeting notes GPT," "support draft GPT"). GPT instructions can override or overshadow your personal defaults for that GPT. Use Custom Instructions for you; use Custom GPTs for the job.

Projects (when your workspace offers project-level instructions): scoped rules for threads inside that project. Best for client voice, brand constraints, or a temporary initiative. Keep global Field 1/2 generic enough that a Project can specialize without fighting you.

Temporary Chat: use it when you want a clean model with personalization off for that session. Useful for A/B testing whether your Custom Instructions help or hurt a prompt, and for topics you do not want tied to Memory.

Common mistakes that empty the benefit

Personality essays in Field 2. "Be witty, warm, world-class, and insightful" burns the budget and gives the model nothing enforceable. Replace adjectives with length, format, and opener rules.

Contradictions. "Detailed explanations" plus "always under 40 words" without a trigger produces random length. Add a scope line: "Short by default; expand only when I say deep dive."

Task workflows in the global boxes. A standing order to "always write LinkedIn posts" will distort coding chats. Move workflows to Custom GPTs or per-chat prompts.

Secrets and client PII in Field 1. Treat the boxes like text that can be retained under your plan's data settings. Keep sensitive material out.

Never testing Temporary Chat. If you only test inside a long thread full of prior corrections, you cannot tell whether Field 2 or the thread history drove the good behavior.

Ignoring presets. A Friendly tone preset plus a Field 2 ban on chit-chat creates mixed signals. Align preset and Field 2, or set the preset to Default and let Field 2 carry the contract.

Updating Field 1 each week with temporary project noise. That churn makes yesterday's context fight today's chat. Put temporary context in the user message or a Project.

Copying a long API system prompt into Field 2. Character limits will truncate it, and middle rules will drop first. Shorten for Custom Instructions; keep the long version for Custom GPTs or the API.

Model notes for ChatGPT in mid-2026

ChatGPT's fast default chat path is often GPT-5.5 Instant. Deeper or slower options in the product lineup may surface as GPT-5.6 Sol (flagship reasoning), Terra, or Luna depending on your plan and picker. Custom Instructions apply across the ChatGPT product surface you enable them on; they are not a substitute for API system messages in your own apps.

On GPT-5.5 Instant, short outcome-first Field 2 text works better than process stacks. Skip "think step by step" as a standing rule. Define what good looks like: answer first, format, fallback when data is missing. On stronger reasoning options, keep Field 2 even shorter and move hard analysis constraints into the user message for that turn.

If you also use Claude Fable 5, Claude Opus 5, Claude Sonnet 5, Gemini 3.5 Flash, or Gemini 3.1 Pro outside ChatGPT, do not expect these Personalization boxes to travel with you. Rebuild the same ideas as project instructions or system prompts in those products. Keep a single notes file with your Field 1 facts and Field 2 rules so ports stay consistent.

Soft next step with PromptMake /text

When Field 2 is a mess of adjectives, paste your rough preference list into PromptMake /text and ask for a compact response contract aimed at ChatGPT. Keep the output under the character budget, strip any leftover sales tone, and paste the cleaned version into Personalization. Use the free daily generations to iterate once or twice, then stop and test inside ChatGPT with the five-prompt suite above.

Custom Instructions remain the control surface. PromptMake only helps you draft clearer text for that surface. After you save, the scoreboard is your own chats.

FAQ

What are ChatGPT Custom Instructions?

ChatGPT Custom Instructions are Personalization settings that store what the model should know about you and how it should answer by default. They apply when Enable customization is on, across plans that expose the feature on web, desktop, and mobile. You edit them in Settings rather than pasting the same preamble into every new chat.

How do I turn on ChatGPT Custom Instructions?

On web or desktop: profile → Settings → Personalization → Custom Instructions, then enable customization and fill the fields. On mobile: Settings → Customize ChatGPT, enable customization, and enter your text. Start a new chat to verify behavior after you save. If nothing changes, confirm the toggle is on and that you are not in Temporary Chat.

What should I put in each Custom Instructions field?

Field 1 holds stable facts about you: role, domain, audience, stack, and standing boundaries. Field 2 holds response rules: opener, length, format, uncertainty handling, and redirects. Keep task-specific workflows out of both fields so a coding chat and a writing chat can share the same account defaults.

How are Custom Instructions different from a system prompt or Custom GPT?

Custom Instructions are the ChatGPT product's account-level personalization fields with a tight character budget. A system prompt in the API or a Custom GPT Instructions box can be longer and job-specific. Use Custom Instructions for personal defaults; use Custom GPTs or API system messages when you need a specialized assistant with knowledge files or tools.

Do Custom Instructions work with ChatGPT Memory?

Yes, they stack when both are enabled. Custom Instructions hold the editable contract you maintain by hand; Memory holds evolving facts the product may recall. Put format and tone rules in Custom Instructions so you can revise them in one place, and use Temporary Chat when you want neither layer for a single session.

Which model do Custom Instructions affect in 2026?

They apply inside ChatGPT for chats that respect personalization, including threads on GPT-5.5 Instant and other models you pick in the product. They do not follow you into third-party apps or raw API calls on their own. For API work, send an explicit system message in your request.

How do I start if the boxes feel blank?

Write five bullets: who you are, who you write for, default length, default format, and one redirect. Turn those bullets into Field 1 and Field 2, staying under the character limit. If wording is the blocker, draft with PromptMake /text, then paste and run the five-prompt test in a fresh ChatGPT thread before you call the setup done.