How to build a Claude Cowork system that runs itself

Executive overview

Most AI tools only read files and give advice — you still do the work. Claude Cowork can read and write files and connected systems on your behalf, enabling fully automated workflows.

Three levels of use build on each other: Do, Make, and Know. Each level increases how much the AI handles without prompting. At the top level, the AI accumulates knowledge across sessions and compounds in value over time.

The shift is from AI as a consultant to AI as an asset that does and remembers the work.

The three levels of Cowork use

1. Do — Give the AI a task; it takes action on your files directly. Ad hoc, prompt-driven.
2. Make — Drop in a single input; the AI reads and writes across multiple connected systems and returns finished outputs.
3. Know — Combines Do and Make, plus a memory file the AI updates each session. Knowledge compounds over time.

Level 1: Do — file organisation example

• Point Cowork at a folder of poorly named files (receipts, contracts, PDFs).
• Prompt specifies: role, task, naming format (date–type–description), subfolder creation, and constraints.
• Key constraint: never delete files, only rename.
• AI logs every rename to a changelog.txt file so it can resume if context fills up.
• System instructions live in claude.md inside the folder — Cowork reads this automatically before every task.
• Once set up, typing "go" is all that's needed.

Level 2: Make — post-meeting processing example

• Drop a meeting transcript into the folder; the AI handles all follow-up without further prompting.
• AI checks Google Calendar for meeting date and attendee emails.
• Simultaneously produces three outputs: a draft follow-up email, an action-items Excel sheet, and an executive summary.
• Subagents (parallel child AI instances) handle each output at the same time — faster, and each gets its own full context window.
• Subagents are only appropriate when tasks are independent of each other.
• Connectors (Gmail, Google Drive, Notion, CRM) must explicitly support write access — check each one in Manage Connectors.

Level 3: Know — compounding memory

• A memory.txt file is created and maintained by the AI across every session in that folder.
• Before each task, the AI reads the memory file; after, it appends new insights.
• Memory entries capture: client communication preferences, recurring themes, key decisions, budget cycles.
• Constraint: the AI never removes previous entries, only adds.
• Over time the AI shifts from a disposable tool to a compounding asset — each session builds on the last.

Setting up system instructions (claude.md)

• Every folder in Cowork can have a claude.md file with persistent instructions.
• Equivalent to custom GPT system prompts, Claude Projects instructions, or Gemini Gems.
• Instructions are read before the AI does anything in that folder.
• For the Know level, bookend the core task instructions with: (1) read memory file first, (2) update memory file last.