Published on [Permalink]
Reading time: 2 minutes

The Setup That Makes Claude Self-improve

I’ve been leaning harder into Claude these past weeks than I ever have before. Real work. Building Felix with Claude Code. Managing multi-step projects with Claude Cowork. (Once only on MacOS, now available for Windows)

What changed? I read Luca Dellanna’s piece on Claude setup and it clicked.

The core insight is deceptively simple: Claude can learn from its mistakes, but only if you set it up that way. Most people use it like Google. Ask a question, get an answer, start fresh next time. No memory, no learning, no compounding improvement.

Dellanna’s approach is different. Configure Claude once to save its mistakes and corrections in files it can access later. You tell it when it’s wrong. It writes that down. Next time, it checks those notes before responding.

This is not rocket science. It’s just good engineering. But the impact is enormous.

I’ve been using this setup with Claude Cowork and the difference is stark. Tasks that used to require constant supervision now run in the background. Multi-step work actually completes. When something breaks, it doesn’t break the same way twice.

Here’s what makes Cowork powerful: skills. Reusable packages of instructions Claude loads when needed. Write a skill once for how you want presentations formatted, and Claude applies it consistently every time.

I’m now building Felix entirely through Claude Code. Not generating snippets I debug for hours. Actual working code. The kind that previously required deep flow state and uninterrupted focus.

The limiting factor has shifted from technical skill to imagination - oh and planning. You still need to have a good grasp on what you want but even there Claude will help prompt you to shape plans.

Three things matter: Claude remembers what you’ve taught it. It runs complex tasks without constant babysitting. It improves with use instead of staying static.

The setup itself is straightforward. You’re just giving Claude access to a folder where it stores notes about what works and what doesn’t. But the implications are substantial. This is productivity infrastructure, not a productivity hack.

Give it a week. You’ll know quickly whether this is noise or signal.

For me, it’s been signal. Clear, compound signal.


Sources:

✍️ Reply by email