A policy says what is allowed, agreements say how you work
An AI policy covers which tools are allowed and which data never goes in. You need one. But a team that uses AI every day runs into different questions: why does the AI answer differently for you than for me, who changed this instruction, and who should have checked this before it went out? That is what working agreements are for.
The five agreements that matter
| Agreement | Why | Where to record it |
|---|---|---|
| 1. Where do shared instructions live? | Otherwise everyone writes their own version in their own chat | A shared project, a CLAUDE.md in the repo, Cursor Project or Team Rules |
| 2. Who maintains them? | An instruction without an owner goes stale | Can edit for one or two people, Can view for the rest |
| 3. What do you always check? | AI writes convincingly, even when it is wrong | In the instructions themselves, plus a human check on anything that goes out |
| 4. Who manages connectors and memory? | A shared credential gives everyone the same access | The Owner or admin of each tool |
| 5. When does something personal get shared? | Good prompts and projects leave with their maker | A standing rule: once two people use it, it moves to a shared place |
To make it concrete, add three agreements you can adopt today: name a reviewer for external output, define when the AI must stop and ask, and review the agreements after an incident or a tool change.
What the tools give you for this
Claude Code. A CLAUDE.md in the project folder (./CLAUDE.md or ./.claude/CLAUDE.md) reaches the whole team through version control. For rules that apply to the whole organization there is a managed CLAUDE.md: on Windows at %ProgramFiles%\ClaudeCode\CLAUDE.md, on macOS at /Library/Application Support/ClaudeCode/CLAUDE.md and on Linux at /etc/claude-code/CLAUDE.md, deployed through MDM or Group Policy. Personal preferences belong in CLAUDE.local.md, not in the shared file.
Cursor. Project Rules in .cursor/rules live in version control. Team Rules apply across the organization and are managed by team admins in the Cursor dashboard, on Team and Enterprise plans. User Rules are personal: no team agreement belongs there. More on this in how to share Cursor context with a team.
Claude (chat). A shared project bundles instructions and knowledge for everyone with access. Owners turn memory on or off under Organization settings > Capabilities. Memory defaults depend on organization creation date, account history and organization type; check both organization and individual settings. On Team, only an Owner adds custom connectors, under Organization settings > Connectors; members then connect on their own. Note that a shared credential in such a connector gives everyone who uses it the same access.
The misconception: an agreement in a document is an agreement with the AI
Many teams write their agreements in a doc or a wiki. A policy document only informs the AI when the tool loads it or retrieves the relevant passages. An agreement only works when it sits where the AI picks it up: the project instructions, the CLAUDE.md, the Cursor rules. Even then it is context, not a lock. The Claude Code docs say CLAUDE.md is not enforced configuration; use hooks for fixed checks and managed settings for centrally managed configuration; instructions in CLAUDE.md stay context and do not guarantee compliance. Keep each file short: Anthropic recommends under 200 lines.
What this means for a team
Every tool has its own place for agreements: Claude has projects and CLAUDE.md, Cursor has its rules, ChatGPT has something else again. Tools have different instruction mechanisms, but some can share an AGENTS.md; keep tool-specific files small and assign one owner to the common rules. At minimum, agree which version leads and who updates the others. See also what happens when your AI champion leaves.
Where Bombos fits
Bombos is a plugin that gives Claude, Claude Code, ChatGPT, Copilot, Cursor and your own agents one shared memory. An agreement is maintained once and made available to connected tools; rules the team teaches the AI along the way get added, traceable, so you can see what an answer rests on. One person can try it first for their own agents.
Related questions
What is the difference between an AI policy and working agreements?
An AI policy says which tools are allowed and which data never goes in. Working agreements say how the team works with AI day to day: where instructions live, who maintains them and what gets checked. You need both.
Is a rule in CLAUDE.md enforced?
No. The Claude Code docs describe CLAUDE.md as context, not enforced configuration. Use hooks for fixed checks and managed settings for centrally managed configuration; instructions in CLAUDE.md stay context and do not guarantee compliance.
How long should an instruction file be?
Anthropic recommends keeping each CLAUDE.md under 200 lines. Short and concrete beats a handbook.
