ChatGPT Custom Instructions vs System Prompts: What to Put Where

ChatGPT custom instructions and system prompts both shape how the model answers, but they live on different levels. Custom instructions follow you everywhere; system prompts shape one task. Most people stuff the wrong one - this guide draws the line, with prompts that already follow the rule.

Custom Instructions and System Prompts Are Not the Same Thing

Every prompt you write ends up somewhere in the model's context, but ChatGPT does not treat all instructions equally. The settings panel, the chat window and a custom GPT each hold their own layer of guidance, and each layer has a different reach. Custom instructions are the global layer: they apply to every chat you start, on every device, until you edit them. A system prompt is the task layer: it applies to one conversation, one GPT or one project. Confusing the two produces bloated prompts, contradictions and mediocre output. Once you see the difference, most prompt problems become placement problems.

What ChatGPT Custom Instructions Do (and Where They Live)

Custom instructions let you share anything you want ChatGPT to consider in its answers, and they apply immediately to all of your chats. OpenAI built them as a persistent profile: you set them once and the model carries them into every new conversation. They work on the free and paid plans and across web, desktop, iOS and Android, so the same style follows you no matter where you open ChatGPT.

The official documentation describes them as a way to avoid repeating yourself in every chat. Instead of typing "remember I work with fashion brands" at the start of each conversation, you store it once. That is the whole point: stable facts and preferences that never change belong in the settings, not in your daily prompts. See the OpenAI help article on custom instructions for the official description.

The Two Fields: About You and How to Respond

Custom instructions are split into two fields, and each one has a job. The first asks what you would like ChatGPT to know about you to provide better responses: your role, your industry, your audience and any context that does not change. The second asks how you would like ChatGPT to respond: tone, format, length and language. If you are a freelance designer who writes in Spanish, the first field holds "I am a freelance designer working with fashion brands"; the second holds "write in a clear, direct tone". Facts go up, style goes down, and the model applies both everywhere.

The 1,500 Character Budget Per Field

Each field has a hard limit of 1,500 characters, roughly 250 to 300 words. That sounds small until you realize what belongs there: only what is stable. If you try to cram task instructions into the settings, you burn the budget on rules that only apply to one kind of chat. A 1,500-character field is plenty when you store identity and style, and it forces you to keep the global layer clean.

Custom Instructions vs Memory

Custom instructions are static: they say exactly what you typed until you edit them. ChatGPT Memory is different. It learns on its own from your conversations and updates over time without you writing anything. The rule of thumb: if it is a stable fact, put it in custom instructions; if it is temporary context from a project, let Memory capture it or write it in the chat. Static settings give you control, Memory gives you convenience, and mixing them up creates duplication.

What a System Prompt Is (in ChatGPT and in the API)

A system prompt is the instruction block that defines the model's behavior for a single run. In the OpenAI API, it is the message with the system role that sets the persona and rules for that specific call. In the ChatGPT interface, it lives inside a custom GPT (the instructions of the GPT) or it is written as the first instruction of the chat. Where custom instructions answer "who am I everywhere?", a system prompt answers "what should this specific task do?"

System Prompts Inside Custom GPTs

A custom GPT is a specialized assistant with its own instructions, optionally its own knowledge files and tools such as search, DALL-E or the code interpreter. The system prompt inside a GPT is the engine of that assistant: it defines the role, the process and the output rules. The practical split is simple: use custom instructions for your personal defaults, and use GPTs for workflows you repeat. One is your voice, the other is a machine you build.

Project Instructions Override Global Settings

ChatGPT Projects add another layer, and it matters more than most people think. Project instructions override your global custom instructions inside that project. If you set "always reply in Spanish" in settings but a project says "reply in English", the project wins. Also, chats started inside a custom GPT cannot be moved into a project. The lesson: if you believe "everything I put in settings applies everywhere", a project is the exception. Local instructions beat global ones when they conflict.

What to Put Where: The Decision Rule

Here is the core rule of this guide, and it fits in one question: does this instruction apply to all of my conversations? If yes, it belongs in custom instructions. If it applies to one task or one workflow, it belongs in the chat, the GPT or the project. "I am a freelance designer working with fashion brands" is global. "For this article, structure the answer with H2 and H3 headings and keep it under 1,200 words" is local. Ask the question, place the instruction, and stop guessing.

Put in Custom Instructions: Stable Facts and Style Preferences

Custom instructions hold everything that is true today and will still be true next month: who you are, what you do, who you write for, and the tone you always want. These are the facts you would otherwise retype in every chat. When a rule survives across projects, it is a candidate for the global layer.

Put in the Chat or GPT: Task-Specific Instructions

Anything tied to a single deliverable belongs in the task layer. The format of one report, the outline of one article, the steps of one workflow: these change constantly, so they cannot live in a static profile. Keep them in the chat prompt or inside the GPT that runs that workflow. This is also where the task-specific prompts from the catalog do their job, because they encode a role, a process and output rules for one kind of work.

Never Duplicate Instructions

When the same rule appears in custom instructions and in a chat prompt, the model weighs both and they often contradict each other. Duplication also wastes your character budget and makes debugging painful: which version is the model following? Pick one layer per rule. If it is global, remove it from the chat. If it is local, remove it from the settings. One source of truth keeps every layer coherent.

Three System Prompts That Follow the Rule

The cleanest way to see the difference is to look at well-built system prompts. Each one below defines a role, a process and output rules for a single task - exactly what does not belong in custom instructions. The SEO Article Writer System gives the model a senior content strategist role plus a numbered process: keywords, outline, write, E-E-A-T, finalize. The YouTube Script Factory repeats the pattern on another domain: role, mandatory structure (hook, intro, beats, outro) and format rules. The Online Course Creator System is a larger system that builds a course in phases - curriculum, scripts, slides, worksheets, launch - and it demonstrates why a task prompt cannot fit inside 1,500 characters of custom instructions. These are task machines, not global settings, and that is exactly why they work.

Common Mistakes to Avoid

Four mistakes show up again and again. First, stuffing task instructions into custom instructions: they apply to chats where they do not belong and they waste the 1,500-character budget. Second, duplicating the same rule in settings and in the prompt, which creates contradictions the model has to weigh. Third, never reviewing custom instructions when your work or style changes, so the profile quietly becomes stale. Fourth, expecting custom instructions to replace a good system prompt: they are complementary layers, not substitutes. A solid setup stores identity and tone globally, runs each task with a dedicated system prompt, and keeps both layers free of duplicates.

Start by opening your custom instructions and deleting every rule that belongs to a single task. Then, the next time you write a prompt, ask the one question from this guide: does this apply to all of my conversations? If it does not, it is not a setting. It is a system prompt.