How to Make Claude Follow Your Rules

Instructions typed into a chat do not survive the session, and lose ground to everything else as the conversation grows. Deliver them over MCP instead and every session Claude starts in your project already has them — whether you remember to mention them or not.

Three steps

1. Connect Contextium to Claude

Add the Contextium MCP server to Claude Code or Claude Desktop. Your workspace — libraries, skills, and rules — becomes readable by the assistant without pasting anything.

2. Write your rules once

Whatever you type at the start of every session is your first ruleset. Give each rule a type: must, must not, prefer, or consult. The type decides whether Claude complies unconditionally, works around a genuine exception, or goes and reads a document first.

3. Attach, then verify

Attach the ruleset to your project and open "See all rules in effect" to confirm what actually reached the assistant — including anything not applying, and why.

For the things Claude must never do

Put it in the workspace baseline as a must not. Baseline rules reach every project automatically and cannot be detached — and a baseline prohibition cannot be overridden by anything.

That asymmetry is deliberate. A project can reasonably override a convention like “use British English”. It should not be able to override “never log customer data”. The baseline is capped at twenty rules so it stays the place for things that genuinely are not negotiable.

Already using a CLAUDE.md file?

It works well for one repository and one person. Rulesets are for when the same standards have to hold across several projects and several people — one place to edit, prohibitions no project can override, and a history of who changed what.

Compare the two

Common questions

Why does Claude ignore instructions from earlier in the conversation?

Chat instructions compete with everything else and lose prominence as the conversation grows. Something said forty messages ago is still present but no longer weighed the way it was. Rules delivered structurally with each request do not decay like that.

Does this work with Claude Code and Claude Desktop?

Yes — Contextium exposes your workspace over MCP, so anything MCP-capable can read your rules. There is also a CLI if you would rather pull context in a terminal or a script.

How do I stop Claude doing something permanently?

Write it as a "must not" in your workspace baseline. Baseline rules reach every project, cannot be detached, and a baseline prohibition cannot be overridden by any project-level rule.

Can I check which rules Claude actually received?

Yes. The rules-in-effect view lists every rule reaching your assistants and, just as usefully, everything that is not applying and why — duplicates, superseded rules, archived rulesets, permissions.

Rules that arrive with every request

Start free