Instructions & Memory in Claude Cowork
You can't edit Claude's system prompt, but there are several documented places to put standing instructions. Each has a different scope, from every conversation down to one project or one folder. Choosing the narrowest one that works keeps Claude focused and saves usage.
The earlier version described workspace-level system prompts, "personas" and precedence rules for a workspace product that doesn't exist. Everything below comes from Anthropic's Help Center and docs.
Where instructions go
| Place | Applies to | Where to set it |
|---|---|---|
| Global instructions | Every Cowork task (every conversation, in the new Claude experience) | Settings → Cowork → Global instructions; in the new experience, Instructions for Claude in Settings → General |
| Project instructions | Every task in one project | The project's settings |
| Folder instructions | Tasks using a connected local folder (desktop) | Added when you select a folder; Claude can update them during a session |
| Skills | Only the tasks a skill is relevant to | Customize → Skills (guide) |
Memory
- Cowork in the cloud shares your chat memory: Claude starts from what it knows about how you work, and what comes up in a task carries back to chat.
- Project memory is scoped to the project and doesn't leak into others.
- By default, Claude doesn't store sensitive topics such as health, race, religion or politics in memory. You can opt in with Include sensitive topics in memory in Settings → Memory.
What to write
- Be concrete. "Deliver reports as a one-page summary followed by a table" works; "be professional" doesn't.
- Put procedures in skills, not instructions. A multi-step routine you repeat, like a weekly report format, belongs in a skill that loads only when relevant.
- Keep it short. Instructions load into every task and count against your usage.
- Say what not to touch. "Never send email without showing me the draft" is worth writing down, but back it with the Manual approval mode. An instruction isn't a guarantee.
In Claude Code
Claude Code's equivalents are CLAUDE.md files (managed, ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and CLAUDE.local.md) and, since v2.1.277, a project's AGENTS.md. Auto memory holds notes Claude writes itself. See What is a CLAUDE.md file?
Sources: the Claude Help Center — Claude Cowork, Claude Code memory docs. Checked 2026-10-09.
← Back to Claude Cowork hub · See also: Projects & Artifacts · Workflow Builder Guide