How do you set up context engineering for a team using AI tools?

Short answer

Set it up in layers. Organization-wide rules in one managed file, project rules in a short file in the repository, personal preferences in a personal file, and team knowledge in a shared place every tool can search. Keep instruction files short and consistent, assign an owner to each layer, and save approved corrections where colleagues' AI tools can retrieve them.

Start from the layers

Claude Code documents separate instruction layers for the organization, the project and the person, which makes it a useful example even if your team also uses other assistants. Its documentation says each session starts with a fresh context window, and two things carry knowledge across: CLAUDE.md files that people write, and an auto memory that Claude writes itself (the first 200 lines or 25KB of it load at the start).

The CLAUDE.md files come in layers:

Layer Where it lives Who it is for Typical content
Organization Managed policy file: %ProgramFiles%\ClaudeCode\CLAUDE.md (Windows), /Library/Application Support/ClaudeCode/CLAUDE.md (macOS), /etc/claude-code/CLAUDE.md (Linux) Everyone in the company Security rules, what never leaves the company, house style
Project ./CLAUDE.md or ./.claude/CLAUDE.md, shared through version control Everyone working on this repo Build and test commands, architecture, conventions
Topic or path .claude/rules/ Work on a specific part Rules for the API folder, for migrations, for docs
Personal, all projects ~/.claude/CLAUDE.md One person Language, tone, personal shortcuts
Personal, one project ./CLAUDE.local.md One person in one repo Local paths, private test accounts

Set it up in six steps

  1. Write the organization layer first, and keep it short. A handful of rules that hold for everyone. Anything that must be enforced, not just suggested, goes in a hook (PreToolUse) rather than in text.
  2. Give every repository a project file. Under about 200 lines. Commands, structure, conventions. If you have more, split it into .claude/rules/ by topic or path. Imports with @path keep files tidy but do not save context: imported text still loads.
  3. Remove contradictions. The documentation warns that conflicting rules may be picked arbitrarily. One rule, one place.
  4. Add AGENTS.md if people use other coding tools. Claude Code reads AGENTS.md by default only when there is no CLAUDE.md. If you have both, put @AGENTS.md in your CLAUDE.md or configure Claude Code to load both, so shared instructions do not have to be duplicated per tool.
  5. Move knowledge out of instructions. Client details, decisions, product facts and process descriptions do not belong in a file that loads every session. Put them where the AI can search them and retrieve the part it needs, the just-in-time approach Anthropic recommends for agents.
  6. Close the loop. Decide how a correction gets back into the shared layer. When someone tells the AI “we never do it that way”, the owner of that layer records it as a rule with its source. Then test it: two colleagues using different tools ask the same question, check which sources each tool received, and confirm that restricted documents stay out of reach.

What goes into the knowledge layer itself (voice, customers, processes, decisions, systems, permissions) is covered in how to give AI agents your company context.

The common misunderstanding

Most teams treat context as a writing task: write a great instructions file, done. Two things go wrong. The file grows, because every incident adds a line, until it is too long to be followed well. And it covers only one tool, while half the team works in ChatGPT, Copilot or Cursor, each with its own memory that nobody else sees.

Context engineering for a team is maintenance, not a document. Short loaded instructions, a searchable knowledge layer, and a habit of feeding corrections back. Longer is not better here either: Chroma’s research on context rot found that models in its tests got less reliable as input grew, and that distracting passages resembling the right answer added errors. The effect differed per model.

What it means in practice

  • Personal memory does not scale. Auto memory and chat memory are useful for one person and invisible to the rest. If a colleague leaves, their tuned AI leaves with them.
  • Tool switching resets context. Instruction files in a repo travel; chat history and built-in memory do not.
  • Someone has to own it. Every layer needs a name next to it, or it decays.

The background on why this matters is in what is context engineering.

Where Bombos fits

Bombos is the shared layer in step 5 and 6: a plugin that gives every AI tool the team uses, from Claude and Claude Code to ChatGPT, Copilot, Grok, Codex, Cursor and your own agents, one memory with team knowledge, rules learned from corrections, and what already lives in mail, documents and repos. Your instruction files stay short; the knowledge behind them is retrieved per question, and you can trace what each answer rests on. It runs in the cloud or on your own server, and one person can start with it for their own agents and bring the team along when it works.

Related questions

Should we just put everything in one big instructions file?

No. Claude Code's documentation advises keeping each CLAUDE.md under about 200 lines, because longer files use more context and are followed less reliably. Knowledge belongs in a place the AI searches, not in a file it always loads.

Can instruction files enforce rules?

No. Claude Code's documentation is explicit that CLAUDE.md is context, not enforced configuration. To block something for real, use a hook that runs before a tool call (PreToolUse).

Does this work if the team uses different AI tools?

Partly. Claude Code can read AGENTS.md, which several other coding tools use. If the repository also has a CLAUDE.md, Claude Code reads only that by default, so import AGENTS.md with @AGENTS.md. Chat history and built-in memory stay inside each tool, so shared knowledge needs a layer outside them.

Who should own the team's context?

One owner per layer: an admin or IT lead for organization rules, the tech lead or team lead per project, and every person for their own preferences. Without an owner, files grow and contradict each other.

Sources

  1. Claude Code docs: How Claude remembers your project
  2. Anthropic: Effective context engineering for AI agents
  3. Chroma Technical Report: Context Rot
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.