How to Build A Self-Improving System with Claude

Summary

The video presents a five-step B.U.I.L.D. framework for creating a self-improving system with Claude Code. The core idea is to give Claude a structured knowledge base, continuously feed it useful data, and use an improvement process that combines automation with human oversight.

1. Base — Create the foundation

• Create a project with a raw folder for original materials and a wiki folder containing organized, linked notes that help Claude find relevant information.
• Document the setup in CLAUDE.md so Claude can follow the project structure.
• Turn repeated tasks into skills—reusable instructions for how Claude should perform them. For example, an “add a new resource” skill could store a file and update the wiki.

2. Upload — Bring in existing data

• Gather past Claude conversations, personal files, email exports, and information about your background and goals.
• Use Claude to analyze these materials and suggest useful skills or system improvements.
• Be mindful of privacy when adding personal data.

3. Inflow — Keep new data coming in

Build repeatable pipelines to collect and process new information. Examples include:

• Claude session history, synchronized into the project.
• Personal work data, such as meeting transcripts, Slack messages, and YouTube captions.
• Curated external content, such as newsletters, books, blogs, or videos. The presenter recommends selecting valuable sources rather than collecting everything.
• Periodic personal reflections, dictated with voice-to-text tools and added as new resources.

Create and test a skill for each pipeline. Mentioned tools and sources include Granola and its MCP for meeting transcripts, Google Takeout or Outlook exports, and Hex or Whisper Flow for dictation.

4. Loop — Review and improve the system

Use a “system improvement” skill to analyze incoming data and recommend changes. Divide recommendations into three groups:

• Auto-approval: Low-risk fixes, such as cleaning data or repairing missed connections.
• Requires confirmation: Significant changes, such as editing or creating skills.
• Needs more context: Items Claude cannot decide without additional information.

Keep proposed changes in a dated Markdown file for review. This middle-ground approach aims to avoid both unchecked automation and the burden of approving every small change.

In Claude Code’s desktop app, create local scheduled procedures that call existing skills. The example setup runs data ingestion on Tuesdays and Fridays at 9 a.m., then runs system improvement later on those days. A separate human-review step lets the user approve or reject proposed changes.

5. Drive — Put the system into use

• Start gradually, remove components that are not useful, and improve skills based on real use.
• Use feedback to improve skills. For example, ask Claude to revise a skill after a poor result.
• Avoid over-engineering. The guiding principle is that action produces information, so experimenting is more useful than endlessly planning.

Guides and tutorials covered

• Prompts for setting up the project structure, knowledge base, and routine skills.
• Prompts for analyzing historical Claude sessions and bulk-loading personal data.
• Guidance for creating session-sync, ecosystem-sync, curated-content-sync, and system-improvement skills.
• A walkthrough of scheduling local procedures in Claude Code and connecting them to reusable skills.

Main speaker and sources

• Main speaker: Austin Marchese, the video’s presenter.
• Sources and people mentioned: Andrej Karpathy and the Anthropic team are cited as influences on the knowledge-base approach. Brian Armstrong is quoted on the value of action.