TL;DR: Every major AI platform gives you two places to put instructions: account-level settings that apply to every conversation, and workspace-level settings scoped to a specific project or notebook. Most users fill in one and ignore the other. Getting both layers right is what separates an assistant that consistently performs from one you constantly correct.
In a previous article on Claude’s profile settings , the focus was on the baseline: how to write account-level personalization that eliminates the need to re-explain yourself every session. That setup still matters, and if you have not done it, start there. But there is a second layer most people ignore entirely: the instructions fields built into Claude Projects, ChatGPT Projects, and Gemini Notebooks. This article covers both layers together, applies the same thinking across all four major platforms, and gives you a framework for keeping everything current as your work evolves.
What Is the Difference Between Account Settings and Project Instructions?
Account-level instructions are always on. Whatever you put there applies to every conversation you have on that platform, whether you are inside a project or not. Think of it as the standing briefing the platform reads before every session: who you are, how you communicate, what you always want and never want.
Project instructions (or workspace instructions, or notebook instructions, depending on the platform) are scoped. They apply only within a specific project or notebook, and they layer on top of your account settings rather than replacing them. The project layer is where you add context that would feel strange or irrelevant in an unrelated conversation: the audience for a particular client, the coding conventions for a specific codebase, the tone rules for one publication.
The practical distinction is this: if an instruction would make sense in any conversation you have, it belongs at the account level. If it only makes sense when you are doing a specific kind of work, it belongs in the project.
How Much Space Do You Actually Have?
Character limits vary by platform and change without much notice. The table below reflects current limits as of this writing. If your instructions are being silently cut off, shorten them and test.
| Platform | Field | Current Limit |
|---|---|---|
| Claude (claude.ai / Claude Desktop) | Single custom instructions field | ~2,000 characters |
| ChatGPT (paid plans: Plus, Pro, Business, Enterprise) | Per instructions field | 5,000 characters |
| ChatGPT (free and Go plans) | Per instructions field | 1,500 characters |
| Gemini | Single instructions field | ~2,000 characters |
| Microsoft Copilot | Single instructions field | ~1,500 characters |
ChatGPT expanded its limit significantly for paid accounts in July 2026, raising each field from 1,500 to 5,000 characters . The instinct to fill that space is understandable. Resist it. A 4,800-character instruction block is harder for the model to follow reliably than a focused 1,200-character one. More space is useful for specificity, not volume.
What Belongs in Your Account-Level Instructions?
The account-level field is your permanent context layer. It should cover four things: who you are and what you do, how you want the assistant to behave, what you always want in a response, and what you never want.
Here is a worked example built for a management consultant. Note the structure: sections are labeled, language is direct, and the total runs under 1,600 characters to stay safely within every platform’s limit.
## INSTRUCTIONS
I am [YOUR NAME], [YOUR TITLE] at [COMPANY NAME], a B2B management
consulting firm (~260 employees, Chicago HQ). I oversee capacity planning,
vendor management, process design, and executive reporting.
### TONE
Direct and professional. Never open with "Great question!", "Certainly!", or
similar affirmations. Start with the answer or the output. Conversational
where context allows.
### FORMAT
- Short tasks (emails, quick answers): no headers, under 200 words.
- Medium tasks (memos, briefs): use headers, 300-600 words.
- Long tasks (reports, analysis): lead with an executive summary, then
structured sections. Use tables for comparative data.
### LANGUAGE
Active voice always. Avoid "leverage," "synergy," "bandwidth" (for people),
"seamless," "deep dive," and any jargon that avoids saying something specific.
Spell out acronyms on first use.
### WHEN UNCERTAIN
Flag uncertain facts with [VERIFY] and continue. Do not present guesses as
facts. Do not invent citations or statistics.
### STANDING RULES
- Do not generate legal or financial advice; flag those questions.
- Make a reasonable assumption, state it, and proceed rather than asking
multiple clarifying questions first.
- Never feign emotions or use language designed to create dependency.
- The final decision always rests with me.
If you want a ready-to-use template for this and a full three-tier reference kit, the AI Reference Kit on GitHub covers the complete setup, including files for your role, your organization, and your active projects.
What to Leave Out of Account-Level Instructions
General instructions are not the place for the following. These belong in your project or notebook instructions.
| Content Type | Why It Belongs in Projects |
|---|---|
| References to uploaded files (“see brand-voice-guide.md”) | Account settings have no files attached, so the reference is meaningless |
| Project-specific context (Project Alpha, Q3 staffing cycle) | Too narrow; it will feel irrelevant in unrelated conversations |
| Detailed output templates (full memo structure, dashboard format) | Clutters general conversations where the format does not apply |
| Stakeholder communication notes | Relevant only when those people are involved |
| Glossaries and terminology rules | Belongs in the workspace alongside the relevant reference file |
The test is simple: would this instruction feel strange in an unrelated conversation? If yes, it belongs in a project.
How Do You Set Up Project Instructions?
Every major platform now offers a workspace layer where you can scope instructions to a specific kind of work.
On Claude (Projects or Cowork), project instructions live in the project’s settings panel. They stack on top of your account preferences, with the project taking precedence on any conflict. You can also upload files directly to the project and reference them in the instructions by name, which is where the real power is. Support documentation is available at support.claude.com .
On ChatGPT (Projects), the instructions field works the same way and you can build a library of uploaded files the assistant draws from. OpenAI’s documentation covers setup at help.openai.com .
On Gemini (Notebooks), instructions are set per notebook. The structure is slightly different in that you can attach sources directly and the assistant retrieves from them inline. Google’s guide is at support.google.com .
Regardless of the platform, a project instruction block should cover six things: the scope of this workspace and what you use it for, the files you have uploaded and when the assistant should use each one, the standing instructions specific to this work, an example of the kind of output you want, what the assistant should do when it is uncertain, and any platform-specific capability settings.
Naming your uploaded files and explaining when to reach for each, is the one most people skip. Without it, the assistant may retrieve from your files inconsistently or not at all. A line like this makes the difference:
brand-voice-guide.md
Contains: Tone rules and annotated writing examples for client-facing work.
Use it: Before drafting any external communication or report.
How Do You Keep Instructions From Going Stale?
Your role, your priorities, and your communication preferences will change. Instructions that are not maintained become noise the assistant learns to work around. A regular review schedule prevents that.
For account-level instructions, a 90-day review is enough. Block 10 minutes in your calendar. Read through each instruction and ask three questions: Is this still accurate? Is this still how I want the assistant to behave in every conversation? Is anything missing that has become a consistent preference?
For project instructions, the review should happen whenever the project itself changes in a meaningful way: a new phase, a change in stakeholders, a shift in output format. Project file uploads should be updated on the same schedule. If a project brief has not been touched in six months, either update it or move it out.
A useful convention is to date your project briefs and mark decisions as active or superseded rather than deleting them. An assistant that can see a decision was reversed is more useful than one working from a log that only shows the current state.
Key Takeaways
- Account-level instructions are always on and should cover your role, tone, format preferences, and standing rules. They should fit comfortably within any platform’s character limit.
- Project instructions layer on top of account settings and should cover only what is relevant to that specific workspace, including the names and usage rules for any uploaded reference files.
- ChatGPT expanded its per-field limit to 5,000 characters for paid users in July 2026, but more space is not an invitation to add more content: it is an invitation to add more specificity.
- The single most underused feature in project instructions is the file mapping: naming each uploaded document and telling the assistant exactly when to use it.
- Set a 90-day recurring calendar reminder to review your account-level instructions, and review project instructions whenever the project scope changes.
Frequently Asked Questions
Do project instructions replace my account-level instructions? No. Project instructions layer on top of your account settings. When you are inside a project, both sets apply, with the project taking precedence if there is a conflict. Outside a project, only your account settings apply.
What happens if I exceed the character limit for custom instructions? Most platforms silently truncate whatever exceeds the limit, which means the end of your instructions, often the most specific and actionable parts, simply disappears. Keep a copy of your instructions in a plain text file so you can verify what actually made it in by pasting and checking the length.
Should I use the same instructions across Claude, ChatGPT, and Gemini? A shared base makes sense for tone and format preferences, since those apply everywhere. But you will likely want to adjust how you reference uploaded files and platform-specific features for each workspace. The AI Reference Kit provides a starting structure that adapts to all three platforms.
How do I know if my instructions are actually being followed? Ask directly in a fresh conversation within the application: “What do you know about how I prefer to receive responses?” A well-configured assistant should be able to summarize your tone, format, and standing rules back to you accurately. If it cannot, your instructions either were not saved correctly or are not specific enough to register.