AI Context: Where to Set It at Account vs Project Level

A modern AI assistant does not offer just one place to write a standing instruction: it offers three. Account, project, prompt. Each has its own lifespan. Here is where to find these settings across five tools, and how to decide what goes where.

Account, project, prompt: three lifespans, not three synonyms

The account holds what is true in every conversation, regardless of topic: your job, your language, how you want answers phrased. The project or workspace holds what is true for one specific case, and only that case: a client's documents, a product's vocabulary, a constraint that only matters this quarter. The prompt, finally, holds what is only true once: today's request.

The most common mistake is writing everything at prompt level. It works, but it means retyping everything in every message, and it stops the assistant from building on a case across conversations. The opposite mistake exists too: putting something in the account that only applies to one project, which then bleeds into an unrelated conversation the following week.

ChatGPT: Custom Instructions and Projects

OpenAI documents the distinction directly in its help centre. Custom Instructions (an account-level setting, found in the personalisation menu) apply to every conversation. A Project is a dedicated space that groups conversations, files and instructions specific to that project; project instructions apply on top of account instructions, and take priority in case of conflict.

The Custom Instructions field now accepts up to 5,000 characters on Plus, Pro, Enterprise, Business and Education accounts, up from 1,500 previously: enough room for a real account context, not just a sentence.

To set up your professional context: open the personalisation settings and describe your job, your sector, the tone you expect in answers. For a project context: create one Project per real case (a client, a campaign), and add the instructions and files specific to that case there.

Claude: Instructions for Claude and Projects

Anthropic distinguishes three layers in its help centre: Instructions for Claude (account-level, under Settings > General > Profile), Styles (tone and level of detail, set per conversation) and Project instructions (per workspace, available on paid plans only). Project instructions layer on top of account instructions; inside a project, both apply, with the project taking priority in case of conflict. Outside a project, only account instructions apply.

This name, Instructions for Claude, replaced Profile Preferences in spring 2026: if an older article mentions "Profile Preferences", it is describing the same setting under its previous name.

To set up your professional context: Settings > General > Profile, write your role and response preferences once. For a project context: create one Project per case; its instructions add to your account instructions rather than replacing them.

Gemini: Personal context and Gems

Gemini separates personal context (Personal context, which replaced the older Saved info setting) from Gems, custom assistants built for a recurring task. Personal context lives in your Gemini account settings and remembers facts that rarely change: your job, your languages, your tone preferences.

One thing to know before relying on it: memory does not apply inside Gems or in Live mode, only in a normal conversation. Availability also depends on account type, age and region: it is confirmed on Google AI Pro and AI Ultra personal accounts in supported regions.

To set up your professional context: account settings, Personal context section, describe your job and preferences. For a project context: create a Gem dedicated to a recurring task, keeping in mind that your account's Personal context does not carry over into it automatically: you need to rewrite whatever is useful directly in the Gem's own instructions.

Grok: Workspaces, memory and Custom Agents

Grok offers three distinct building blocks: a cross-conversation memory (optional, viewable and editable in your data controls), Workspaces that group conversations, files and instructions by project, and Custom Agents, up to four configurable agents with their own name, personality and instructions.

This is the least stable of the five tools on this exact point: the custom-instruction character limit changed several times in 2026 according to the sources consulted for this article. Check the limit shown in your own settings rather than a figure published a few months ago.

To set up your professional context: account memory picks up what recurs across your conversations, with no dedicated setting to fill in ahead of time. For a project context: one Workspace per case, with its own instructions and files.

Mistral: Memories and Projects, under the name Mistral Vibe

Le Chat is now called Mistral Vibe: official documentation states the old name is being phased out, with a six-month transition window for Enterprise organisations. An article or screenshot that still refers to the Mistral "Chat" is describing the same interface under its previous name.

Mistral Vibe separates Memories (information retained about you, your preferences and your professional context, at account level) from Projects (workspaces dedicated to a case). A third layer, the Knowledge Base, is gradually replacing Memories in the unified Work experience, to store a broader range of information.

To set up your professional context: Memories tab under the Intelligence menu, turn memory on and let it learn, or fill it in directly. For a project context: create a Project in the Vibe Work interface.

What never changes, whatever the tool

The principle stays the same everywhere: something true across all your exchanges goes in the account, something that only applies to one case goes in a dedicated project, a one-off request stays in the prompt. What all five tools also share: none of these settings is a protected space under data protection law. Sensitive personal data, about you or about a third party, has no place there, at any level.

Your turn

Two concrete actions, worth doing now rather than just reading about:

Set up your professional account context. On the tool you use most, open the relevant setting above and write out your job, your sector, the tone you expect, once. Every prompt after that gets shorter, with nothing to retype.

Set up your project contexts. For each active case (a client, a campaign, ongoing research), open a dedicated workspace rather than mixing everything into one conversation. The context stays there, from one session to the next.