What to put in CLAUDE.md and what to attach to a single task
Claude Code reads CLAUDE.md from the repository when it starts. Codex and many other agents have the same kind of file, AGENTS.md. It's a handy place for everything the agent should know about the project: how to run the tests, where things are, what the code rules are.
Over time the file grows, and instructions appear in it that aren't needed in every task.
An example: reviewers after development
A useful technique for code tasks is independent reviewers. After development the agent launches separate subagents that didn't see how the code was written. They read the changes and check them against the task and the project rules. One checks whether what was asked for got done. Another looks at style and unnecessary code. The agent fixes what they find before saying "done".
Such an instruction takes about fifteen lines. Put it in CLAUDE.md, and the agent gets it in every session: in a task about landing page copy and in a log investigation, where there is no code to review. That's extra context, and sometimes extra work.
Writing it by hand in every task isn't convenient either. It's easy to forget, and copying it from an old task makes it easy to pick up an outdated version.
An instruction for one kind of task
In Tascue such an instruction becomes an include, for example "Independent reviewers". You check it in tasks where code changes, and when the task is copied it goes to the end of the prompt. Tasks about copy and research simply don't have it.
If the include appears when the backlog already has many tasks, you can add it to all of them at once: "Apply to all backlog tasks". Tasks in progress and done tasks won't change.
What goes where
| Instruction | Where to keep it |
|---|---|
| Needed in any task in this repository: how to run the tests, how the project is built | CLAUDE.md |
| Needed in tasks of one kind: code review, checking in the browser, copy rules | Project include |
| Needed in all projects: "don't commit" | Shared include, included in new tasks by default |
| Needed in one task | Task text |
A simple check for a new instruction: is it needed in every task? If not, it belongs in an include, and CLAUDE.md stays short enough to read in a minute.
The prompt is assembled from the task on its own: intro, fields, text, includes. You can paste it into Claude Code after "Copy" or hand it over through MCP, and the agent gets the same task in full.