How to Give Claude Global Instructions

A global instruction for Claude is one that applies to every conversation rather than just the current message. Claude has three scopes for this: account-level preferences in Claude.ai, project-level instructions in Claude Projects, and user-level rules in Claude Code. Each scope covers a different set of situations. A tool like fenbs adds a fourth layer: team-wide decisions and rules that every connected AI app reads, regardless of which app each person uses.

Understanding which scope to use saves you from repeating the same instruction in every chat. Write the rule once, in the right place, and Claude follows it automatically.

Account preferences in Claude.ai

Claude.ai has a custom instructions field under Preferences. What you write there applies to every new conversation in your account. This is the right place for preferences that are always true for you, regardless of the project: your preferred tone, your language (American English, not British), your name or role, format preferences like "no preamble," and recurring instructions like "always give me the reasoning, not just the conclusion."

Custom instructions have a character limit. Keep them focused on the rules that genuinely apply to everything. If a rule only applies to a certain kind of work, put it in a project instead.

• Language and spelling preference: "Use American English throughout."
• Format rules: "No preamble. Answer directly. No summary at the end."
• Tone: "Write in plain, direct language. No hype."
• Your context: "I am a software developer. Assume technical literacy."
• A standing disagreement rule: "Disagree with me when you have a real reason to."

Project-level instructions in Claude Projects

Claude Projects lets you write a set of instructions that apply to every conversation within that project. These are more specific than account preferences: they describe what the project is for, any constraints unique to that project, and rules that would not make sense outside it.

A project for legal research might include: "Cite the primary source for every legal claim. Use US law unless specified otherwise. Do not give legal advice; describe legal frameworks." A project for content writing might include: "Write in second person. Use short paragraphs. The brand voice is direct and practical."

Project instructions supplement account preferences. Claude reads both. If they conflict, the project instruction usually takes precedence for that project.

User rules in Claude Code

Claude Code has two places for global instructions. The first is a CLAUDE.md or AGENTS.md file at the root of a repository. Everything in that file applies to every Claude Code session in that repository: what tools to use, what files to avoid, code style rules, and anything the assistant should always or never do in this codebase.

The second is a user-level instructions file that applies across all repositories. This is the right place for personal preferences that carry over from project to project: "Always show me the full file before making a change." "Do not assume I want code committed automatically." These instructions follow you rather than living in one repo.

For teams, the CLAUDE.md at the repo root is the shared source. Everyone working in that repo gets the same instructions. Personal preferences go in the user-level file and do not affect teammates.

What belongs in global instructions

A good global instruction is a rule that is always true. It does not depend on the specific task, the current context, or what stage the work is at. Format preferences, language preferences, communication style, and recurring constraints are all good candidates.

A poor global instruction is one that only applies some of the time. "When writing code, always add comments" belongs in a project where you are writing code. It does not belong in account preferences, where it will apply to every conversation including the ones about recipes and marketing copy.

Decisions Claude should always follow

Some global instructions are not preferences. They are decisions that a person made and that hold from now on. "Do not commit code without asking first." "Always confirm the budget has been approved before finalizing a cost estimate." These are rules with a reason behind them, not just style choices.

For personal use, adding these to your Claude.ai preferences or your CLAUDE.md works. For a team, you need a record that explains why the rule exists and who decided it. A rule recorded in fenbs shows the decision, the person who made it, and the reason, so a new teammate or a new AI session understands not just what to do but why.

How to write effective global instructions

Keep global instructions short and specific. Vague instructions like "be helpful" add nothing. Specific instructions like "Do not use the phrase 'certainly'" or "Answer in plain language, not marketing language" change Claude's behavior in a measurable way.

Review your global instructions periodically. Instructions that made sense for one project can become noise when you move on. A global instruction that you no longer need makes every future session slightly harder to manage. Keep the list to rules that are genuinely current and always relevant.

How to do this with fenbs

1. Create a free account and set up a project for the work you want to give standing instructions for.
2. Go to Skills and prompts and add the reusable instruction sets you want every connected assistant to be able to read: format rules, style rules, and recurring prompts.
3. Go to Decisions and rules and record any team-wide rule with the person who decided it and the reason. Connected assistants read these first in their context.
4. Connect your AI apps from Integrations. Claude, ChatGPT, Cursor, Codex CLI and other supported apps connect over MCP with a browser sign-in.
5. At the start of each session, ask the assistant to read your project context. It will find the rules and skills and follow them without you retyping them.