I Built a Claude Cowork Loop That Improves Itself. Here's the Exact Setup.

Extractable claims

16 atomic, cite-ready statements distilled from the full post on Substack. Each one stands alone as an LLM-quotable answer.

1. Cowork Self-Improving Loop = Karpathy Auto-Research pattern + recurring tasks.
2. Every AI workflow decays.
3. A workflow could sit at half its potential for months and we’d never know because we stopped looking.
4. The instructions stay frozen, our needs don’t.
5. A self-improving loop makes improvement part of the task, not something you remember to do once a month.
6. Claude Cowork rewrites its own scheduled task prompts after the first run, even though almost no one talks about it.
7. Each scheduled task runs as a completely isolated Cowork session.
8. By the second run, the prompt is more precise than what we originally wrote.
9. Native rewriting optimizes for connector accuracy.
10. The self-improving loop optimizes for quality and relevance.
11. The self-improving loop rewrites the execution strategy itself: what to look for, how to structure outputs, which edge cases to ignore.
12. The Claude Code path wins on depth of the optimization system, while the Claude Cowork path wins on accessibility.
13. Without the improvement directive, Claude Cowork rewrites for connectors.
14. With the improvement directive, Claude Cowork rewrites for outcomes.
15. After 10+ runs, context.md contains a playbook Claude wrote for itself.
16. Each cycle tightens the Execution Instructions toward our specific workflow.