Guide
Stop re-explaining your project to Claude Code and Cursor
Every new session, you type the same paragraph: what the project is, what's half done, what you decided not to do. Here's how to write it down once where Claude Code, Codex and Cursor read it, and keep it current without doing that by hand.
What a session starts with
A coding agent starts each session with its instruction file, any notes it saved, and the code. The code shows what exists, not why. The why, and what's next, is the part you keep retyping.
1. Rules that don't change
Put the stable things in the instruction file: how to build and test, conventions, what not to touch. That's CLAUDE.md for Claude Code and AGENTS.md for Codex, Cursor and most others (which one you need). Keep it short: a long file gets skimmed, by agents too.
Claude Code also keeps its own notes between sessions with auto memory, which you can review with /memory. Cursor uses project rules. Both are worth checking now and then, because they keep old notes after they stop being true.
2. Where things stand
The part that changes needs its own place. A short section the agent reads first and updates when it's done works well:
## Where things stand
- Read this before starting. Update it when you finish a task.
- Current goal: ...
- Decided: ... (why)
- Tried and dropped: ... (why)
- Next: ...“Tried and dropped” saves the most time. Without it, the next session suggests the approach you already threw out.
3. Resume when it's the same task
In Claude Code, claude --continue picks up the last session and --resume lets you choose one. That's right for finishing the same piece of work. For a new task, start fresh and let the files above carry what matters. When a long session gets heavy: /compact vs /clear.
4. What was decided outside the editor
The rest of what you explain happened somewhere else: the pricing agreed on a call, the bug a customer reported by email, the approach you settled in a different agent last week. None of it is in the repository.
Connect Cachejay and the agent can look it up. It keeps your calls, what you read and wrote, and your Claude Code and Codex sessions, so a decision from last week's session in one agent shows up in another today. Choose Connect AI Apps in Cachejay's menu and pick Claude Code, Codex or Cursor. Agents also report back what they finished, so where things stand stays current without you writing it.
To stop them asking questions you've already answered: stop Claude Code and Codex asking what you already know. For memory across sessions in more depth: how to make Claude Code remember between sessions.