Table of contents
- 1. The "applies to every chat" fields at a glance
- 2. Custom Instructions in ChatGPT
- 3. "Instructions for Claude" in Claude
- 4. Custom instructions in Gemini
- 5. What to write: three principles from an official guide, plus an example
- 6. What not to write
- 7. Checklist for when they don't work
- Summary
- FAQ
If you are tired of typing "lead with the conclusion" or "keep it in plain English" every time, you can write it once in an instruction field that automatically applies to every conversation. ChatGPT, Claude and Gemini all have one.
The three, however, call the field different things, put it in different places and let you write different amounts. On top of that, their official help pages spell out situations where what you wrote does not take effect: inside projects, inside Gems, on work accounts and in temporary chats.
This article lays out where each field lives and its character limit, as checked against each company's official help, then uses Anthropic's official prompting guide to show how to write instructions that work, with an example. It ends with a checklist of where to look when they don't work.
The same "applies to every chat" field, with different names and limits
What each company's official help says (checked in September 2026)
Limit: 1,500 characters on Free and Go / 5,000 on Plus and up. Inside a project, project instructions take precedence
Limit: not published. On company plans, organization instructions written by admins (up to 3,000 characters) take precedence
Limit: not published. Not available inside Gems or with work and school accounts
Sources: OpenAI Help, Claude Help (organization instructions), Gemini Help
1. The "applies to every chat" fields at a glance
UI names are as shown in English. Where this field lives and how much it holds have changed several times over the past year, so the table notes when it was checked.
| Item | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Name in the UI | Custom Instructions | Instructions for Claude | Instructions for Gemini |
| Location (computer) | Settings → Personalization | Your initials (lower left) → Settings | Settings & help → Personal Intelligence |
| Character limit | 1,500 on Free and Go, 5,000 on Plus, Pro, Business, Enterprise and Education | Not published | Not published |
| Where it applies | All chats, including existing ones | All conversations | All chats |
| Where it doesn't apply or gets overridden | Inside projects, temporary chats (by default) | Anything that conflicts with organization instructions | Inside Gems, work, school and supervised accounts |
Checked against each company's official help in September 2026. Sources are given in each section.
2. Custom Instructions in ChatGPT
Where to find it
According to OpenAI's official help, on a computer (web or desktop app) it is under "Settings" → "Personalization", and in the iOS and Android apps under "Settings" → "Customize ChatGPT". Make sure "Enable customization" is turned on, then write in the "Custom Instructions" field.
Character limit
The same help page says you can save up to 1,500 characters on Free and Go, and up to 5,000 characters on Plus, Pro, Business, Enterprise and Education. According to the official release notes, the limit for Plus and higher plans was raised from 1,500 to 5,000 characters on July 15, 2026. Older articles that say "up to 1,500 characters" predate this change.
Behavior worth knowing
- Edits take effect immediately, even in existing conversations. They used to apply only to new conversations, but the release notes say this changed on November 7, 2025
- Inside a project, project instructions take precedence. The Projects help page says project instructions apply only within that project and override your global custom instructions
- Temporary chats don't use them by default. According to the August 27, 2026 release notes, turning on personalization when you start a temporary chat now lets it use your custom instructions and memory. You can't switch this after the conversation has started
- People who view a conversation through a shared link can't see what your instructions say (official help)
3. "Instructions for Claude" in Claude
Where to find it
According to Claude's official help, click your initials in the lower left corner → "Settings" → "Instructions for Claude". Whatever you write there applies to all conversations.
Character limit
The official help does not state a limit. Some articles online give a specific character count, but the numbers differ from one article to the next, and no official source backing them up could be found. Even if you can write a lot, short, specific instructions work better, as §5 below explains.
Project instructions and organization instructions
- Project instructions are written from "Set project instructions" on the project screen and apply only to chats inside that project (official help). The official help does not say which wins when they conflict with Instructions for Claude
- Organization instructions are set by Owners and Primary Owners on Team and Enterprise plans. The official help says they have a limit of 3,000 characters, and when both they and your personal instructions exist, organization instructions take precedence. Changes can take up to an hour to take effect
- The same help page also says that Claude won't follow instructions that conflict with its core training. The instruction field is not a switch for turning off safety behavior
4. Custom instructions in Gemini
Where to find it
According to Gemini's official help, on a computer it is under "Settings & help" → "Personal Intelligence" → "Instructions for Gemini". The instructions apply to all chats and are also used in the mobile app, in Gemini in Chrome (only in countries where it is available) and on smartwatches.
Character limit
The official help does not say.
Where it isn't available (stated in the official help)
- It isn't available in certain features, like Gems. When you work in a Gem, assume your global instructions don't apply
- The feature isn't available at all with work, school or supervised Google Accounts. That is why you can't find the setting when signed in with a company account
- The help page says instructions like "forget this topic" or "avoid this subject" don't always work perfectly
5. What to write: three principles from an official guide, plus an example
Of the three companies, the most concrete guidance on writing instructions comes from Anthropic's official prompting guide. It is written for people using the API, but its principles for writing instructions apply just as well to the instruction field. It comes down to three points.
The guide's rule of thumb: show your prompt to a colleague with minimal context, and if they would be confused, the AI will be too. "Lead with the conclusion in one or two sentences" beats "make it easy to understand"
Explaining why you want something makes replies more likely to fit your goal. "Put it in a table, because I'll paste it into a spreadsheet" beats "put it in a table"
The guide's example: instead of "Do not use markdown in your response", try "Your response should be composed of smoothly flowing prose paragraphs"
Example: a template you can adapt
The example below follows the three principles above. Delete any lines that don't apply to you, and rewrite the rest to fit your work and what you use AI for.
About me
I build websites. My specialty is design, and I'm a beginner at programming.
How to reply
Start with the conclusion in one or two sentences, then explain the reasons. I often read replies in between other tasks.
When comparing several options, put them in a table.
The first time a technical term comes up, add a short explanation.
How to work
If information is missing, ask me instead of guessing.
If you aren't sure about something, say so.
Language
When I paste text or sources in another language, still reply in English.
The last line guards against the reply language drifting, the same problem covered in why AI suddenly answers in the wrong language. The instruction field is not a hard language lock, though, so when you paste a long text in another language, it is more reliable to state the reply language in the request itself as well.
If you want to learn more about making a single request work, the practical guide to prompt engineering is a good next step. In coding tools such as Claude Code and Codex, the equivalent of these every-conversation instructions is a file, CLAUDE.md or AGENTS.md. Those are covered in our collection of instruction file templates.
6. What not to write
- Secrets such as passwords, API keys or card numbers. The instruction field is loaded into every conversation. ChatGPT's official help says information from your use of custom instructions is used to improve models (and explains how to opt out)
- Assumptions that only matter for one specific job. Materials and rules for a single project are better placed in project instructions than in your global instructions, where they won't get in the way of other conversations
- Negative instructions such as "forget about X". Gemini's help states plainly that instructions to forget or avoid something don't work perfectly. As Principle 3 says, phrase it as what you want instead
- Instructions that contradict each other. Putting "be concise" next to "explain in detail" leaves it to the AI to decide which one wins. Separate them by situation, or pick one
7. Checklist for when they don't work
When what you wrote isn't taking effect, check where you are right now before looking at the wording. Here are the situations the official help pages list as ones where instructions don't apply.
| Where you are | What's happening | What to do |
|---|---|---|
| Inside a ChatGPT project | Project instructions override your global custom instructions | Write the same thing in the project instructions |
| A ChatGPT temporary chat | Custom instructions aren't used by default | Turn on personalization when you start it |
| Inside a Gemini Gem | Instructions for Gemini aren't used | Write it in the Gem instead |
| Signed in to Gemini with a work or school account | The feature isn't available at all | Include it in every request |
| Using Claude on a company Team or Enterprise plan | Where they conflict, organization instructions take precedence | Ask your admin what the organization instructions say |
| None of these (a regular chat) | The instructions are vague or contradict each other | Rewrite them using the three principles in §5, and put important requirements in the request too |
The instruction field is not a setting that forces the AI to comply. It is text passed along as background for every conversation. For the broader problem of AI not following rules, see why AI ignores your rules, and how to fix it.
Summary
| Service | Location | Limit | Watch out for |
|---|---|---|---|
| ChatGPT | Settings → Personalization | 1,500 characters (Free, Go) / 5,000 (Plus and up) | Projects, temporary chats |
| Claude | Settings → Instructions for Claude | Not published (organization instructions: 3,000 characters) | Organization instructions on company plans |
| Gemini | Settings & help → Personal Intelligence | Not published | Gems, work and school accounts |
- All three have an instruction field that applies to every conversation, but the name, location and limit differ for each
- Write instructions that are specific, give the reason, and say what to do instead of what not to do
- Leave out secrets and assumptions that only matter for one job. Put the latter in project instructions
- When they don't work, first check whether you are in a project, a Gem, a temporary chat or a company account
FAQ
Q1. How are custom instructions different from memory?
According to ChatGPT's official Memory FAQ, saved memories work much like custom instructions, but custom instructions are something you write and manage yourself, while memory is updated automatically by the AI based on your conversations. Both are used as background when the AI writes a reply. Rules you want followed are more reliably placed in custom instructions, which don't change on their own.
Q2. If I edit them, do the changes apply to my existing conversations?
For ChatGPT, according to the official release notes, since November 7, 2025 changes apply immediately to all chats, including existing ones. Claude's and Gemini's help pages say the instructions apply to all conversations (chats), but don't specify when a change reaches a conversation already in progress. If you want to be sure, start a new chat.
Q3. I get "error (13)" when I try to save custom instructions in Gemini.
The official Gemini community has many posts reporting error (13). Some reports say it appeared in the middle of an ordinary conversation, not just when saving custom instructions. However, no official explanation of the cause from Google could be found. Try the usual steps and see whether it clears: wait a while and try again, reload the page, or shorten your instructions before saving. For the May 2026 incident in which Google named this error, and the order of steps to try, see our article on Gemini error (13).
Q4. Does this instruction field also apply when using the API?
No. OpenAI's official help says there is no API for custom instructions, and that you should use system messages in the API. In the API and in app development, the standard approach is to write these instructions in the system prompt instead of an instruction field.
