ChatGPT custom instructions are account-wide preferences. Project instructions belong to one project and, according to OpenAI, override conflicting custom instructions. A system prompt is the same idea in a more technical form: standing rules sent before the user's task. If you are trying to keep a book’s voice and limits consistent, the question is not which label sounds smarter. It is which layer should hold which rule.
OpenAI’s Help Center article ChatGPT Custom Instructions says custom instructions let you share anything you want ChatGPT to consider, and that they are applied immediately to all chats. The article Projects in ChatGPT says projects keep related chats, files and instructions together, and that “Project instructions apply only within that project and override your global custom instructions.” The custom-instructions article also notes that there is no custom-instructions API, and that Chat Completions system messages should be used for a similar effect.
For a fuller definition of roles versus system prompts, see Role Prompting vs System Prompts. This article is about where those standing rules live in ChatGPT, and how that compares with a system message. Other apps use different names, and the same wording does not always travel cleanly.
For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up. Standing rules for a book, such as the reader, the voice and the “do not invent sources” rule, belong in a reusable workspace, not only in one chat window. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts and standing-rule drafts written for this article. They are not paid Builder prompts, and they are not recorded model outputs.
Product settings change. The details below follow the OpenAI Help Center pages as they stood when this article was written. If a screen in your account looks different, trust the current Help Center text over this summary.
What belongs in each layer?
A useful split is: durable habits at the account layer, book-specific rules at the project layer, and the current job in the prompt.
| Layer | Scope | Good for | Poor for |
|---|---|---|---|
| Custom instructions | All chats, once enabled | “Be direct. Ask before inventing a figure. Prefer British or American spelling.” | The outline of one book, a client’s confidential notes, or this week’s chapter |
| Project instructions and files | Only that project | The reader, the book’s job, standing constraints, and source files for that book | Preferences you want in every unrelated chat |
| Chat prompt (the task) | This message or thread | The section to write, the notes for this page, the length, the requested format | Repeating your entire style guide every time, if a project can hold it |
| API system or developer message | That API call or conversation | The same standing rules, when you are not in the ChatGPT UI | Assuming the ChatGPT screens and the API are identical |
OpenAI’s custom-instructions article says updates apply immediately across chats, including existing conversations, and that earlier versions can still appear in old chat history. It also states character limits: 1,500 characters for Free and Go users, and 5,000 for Plus, Pro, Enterprise, Business and Education users. If your style guide is longer than that, it cannot all live in custom instructions. Put the long guide in a project file, or keep it in your writing workspace and paste the relevant section.
The same article says information from custom instructions may be used to improve model performance unless you opt out, and it warns that relevant instruction text might be shared with third-party plugins you use. That is a reason not to store unpublished manuscripts, client secrets or private research only in a setting you do not fully control. Keep a copy you own.
How do project instructions and files change the job?
A project is a workspace. OpenAI says you can add instructions, upload files, and keep related chats together so ChatGPT can use that context for ongoing work such as writing or planning. That is a better home for one eBook than custom instructions are. The floor-rate book can have its own instructions and its own notes file. Your travel-journal book can have another. Custom instructions should not try to hold both.
Because project instructions override conflicting custom instructions, a project can tighten or relax a global habit. If your account-wide rule is “keep answers short,” a book project can say “chapter drafts should be 700 to 900 words unless the prompt says otherwise.” The override is useful. It is also a source of confusion if you forget it is there and wonder why a chat outside the project behaves differently.
Files in a project are not a substitute for a clear prompt. If you upload 40 pages of notes, still tell the model which notes apply to this section, and still forbid invented sources. Long context can help, but it can also bury the one figure you care about.
Example 1: write standing rules at the right layer
A weak habit is to paste a full manifesto into every chapter prompt, and also into custom instructions, and also into a project. The three copies then drift.
Account-wide custom instructions can stay short:
Be direct. If a figure, quote or study is missing, ask me or label
it "Need from me." Do not invent sources. Prefer clear headings
and short paragraphs. If my message includes a specific format,
follow that format over these defaults.
Project instructions for one book can be more specific:
This project is for a short eBook for new freelance designers who
undercharge. The book's job is to help them name a floor rate
before the next inquiry, not to build a full pricing menu.
Standing rules:
- Use only notes and files in this project, plus what I paste.
- The only publishable example figure is the $90 floor for a
one-page logo revision, unless I replace it in the prompt.
- Do not add discount policy or package-menu chapters unless I
ask for them.
- End teaching chapters with one action the reader can complete.
The chapter prompt can then be short:
Write the "Name the floor before the call" section, 280 to 320
words, from the outline in this message. Do not change the
standing rules.
If you later move to an API, the project block is what belongs in a system or developer message, with the section request in the user message. The custom-instructions Help Center article is explicit that system messages are the API counterpart, not the ChatGPT setting itself.
Example 2: diagnose a conflict between layers
When a reply ignores a book rule, do not only rewrite the chapter prompt. Check the stack.
I asked you to stay inside the $90 floor example, but the draft
mentions an industry average. Please tell me:
1. Which instruction you treated as highest priority
2. Whether a project instruction, a custom instruction, or my
last message caused the added figure
3. A corrected draft that removes anything not in the project
notes or this chat
If you cannot see my custom instructions or project instructions,
say so, and work only from this message.
The model may not be able to quote those settings back to you reliably. That is another reason to keep your own copy of the rules.
Example 3: do not put a whole book in custom instructions
This overloads the account layer:
You are writing my 12-chapter pricing eBook. Chapter 1 is about...
Chapter 2 is about... Always use these 20 sample emails...
Those lines will leak into unrelated chats, hit the character limit, and go stale when the outline changes. Move them into a project, or into a document you attach for that book, or into a Notion writing assistant setup like the one discussed in System Prompts for a Notion AI Writing Assistant.
How does this fit a workspace you actually own?
Chat tools are convenient. They are a poor single home for the canonical rules of a book. Keep the standing brief in a writing workspace you control. Spec Builder can hold the reader, the job and the constraints that should appear in every chapter brief. Drafter and Polisher then run against those rules. If you also use ChatGPT projects, paste or upload the same brief, and update both places when a rule changes.
In the Prompt Engineering Lab, a useful test is to change only where a rule lives: in the task prompt, or in a standing-rules block you treat as a stand-in for project instructions. Predict whether a conflicting chapter request still obeys the floor-rate constraint, then try a failure case that asks for “industry averages.”
FAQ
What is the difference between ChatGPT custom instructions and projects?
Custom instructions are global. OpenAI says they apply to all chats once they are enabled, and that updates apply immediately. Projects group chats, files and instructions for one body of work. Project instructions apply only in that project and override conflicting custom instructions. Use custom instructions for habits you want everywhere, and a project for the rules and files of one book.
Do project instructions override custom instructions?
Yes, according to OpenAI’s Projects help article: project instructions “override your global custom instructions.” If the two layers conflict, the project rule should win inside that project. Chats outside the project still follow the custom instructions, if they are enabled. If your account’s behavior disagrees with this, check the current Help Center text.
Where should I put standing writing rules for a book?
Put book-specific rules in a project, or in a workspace you own, and put only short, global habits in custom instructions. Put the current section, notes and format in the chat prompt. If you use an API, put the standing rules in a system or developer message. Keep a copy of the rules outside the chat app so you can reuse them in Notion, in another model, or after a product redesign.
Before you hit send
- Is this rule true of all your work, or only of this book?
- If it is book-specific, is it in the project (or your workspace), not in custom instructions?
- Is the task prompt still clear about this section’s job and limits?
- Have you checked for a conflict between layers?
- Do you have a copy of the standing rules you can paste elsewhere?
Standing rules are useful when they are short, current and stored in the right place. They are a mess when every layer holds a different version of the book.