Claude Code keeps losing context. Give it a plan instead of a paragraph.

You know the ritual. New Claude Code session, and before any work happens you type the paragraph you've typed thirty times: what the project is, what stack it's on, what you decided about auth last week, what the agent must not touch. Five minutes of briefing, every session.

It's worse than tedious. It's lossy. You never explain it the same way twice, and the agent that refactored your API yesterday made decisions today's session has never heard of.

Why it happens

Context windows end. Sessions end. /compact summarizes, and summaries drop the exact detail you'll need two hours later. CLAUDE.md helps, but it's a static file. It goes stale, it grows, and it lives in one repo, invisible to the chat window where you think, the Cursor window where a teammate works, and the cloud run going overnight.

The root problem isn't the context window. Your project's state has no home. It's spread across chat histories, half-remembered decisions, and markdown files that were accurate three weeks ago.

Give the state a home

Instead of pushing context into every session by hand, give the project one plan and let every agent pull from it.

• The plan lives in one place. Modules, features, tasks, acceptance criteria, in a workspace you edit and reorganize like a project tool.
• Every agent connects over MCP. Claude Code, Cursor, Codex, or a chat window. The agent reads the current state and starts caught up instead of cold.
• Decisions and docs persist. Rules, memory, and documentation live with the project. An agent that hits an old decision reads it rather than guessing.
• Work reports back. Agents update status, post comments, and flag blockers through the same connection. The plan absorbs what happened in the last session, not just what you told it.

That last point is the one nothing else does. A CLAUDE.md tells the agent where you were when you last edited the file. A live plan tells it where the project is now, including what three other sessions did this morning.

In practice

1. Scope the project. Paste an idea or import the repo, get a structured plan in about a minute, then refine it.
2. Connect Claude Code. The MCP server is hosted, so it's a Connect button or one pasted line. Do the same in Cursor or wherever else you work.
3. Start every session with nothing. The agent pulls the plan, the stack, the decisions, and its next task itself.
4. Watch status and comments update on the board as sessions work, in every editor and in cloud runs you started from your phone.

The briefing ritual doesn't get shorter. It goes away.

"Isn't this a project tool with extra steps?"

It's a project tool with fewer steps, because agents are users of it. Linear and ClickUp organize work for people and go stale the moment an agent touches code. A plan that agents read from and write to stays true because it's the interface the work moves through.

Your context window will always end. Your plan doesn't have to.