Building a Self-Improving System with Claude Code

The B.U.I.L.D. Framework

Austin Marchese outlines a five-step process to create a self-improving system using Claude Code. The framework relies on a structured project directory containing a raw/ folder for source data and a wiki/ folder for indexed references, enforced by a claude.md file that acts as a system prompt. The system utilizes "skills"—reusable prompt-based workflows—to handle repetitive tasks like ingesting new resources or syncing conversation history.

Data Ingestion and Pipelines

The system functions as a data lake that requires constant inflow to remain relevant. Marchese recommends three primary data sources: AI conversation history, personal ecosystem data (emails, Slack, meeting transcripts), and curated external content (newsletters). These are managed via "sync" skills that process raw data into the project's knowledge base. To automate this, the system uses local routines within the Claude Code desktop app to trigger these sync skills on a set schedule, such as Tuesdays and Fridays.

The Improvement Loop

Rather than full automation, which risks "system drift" (where the AI optimizes for the wrong metrics), Marchese advocates for a bucketed approach to improvements:

• Auto-approve: Low-risk tasks like data cleanup or fixing broken links.
• Need signoff: Higher-stakes changes like editing or creating new skills, which are written to an output/review.md file for human approval.
• More context required: Items that the system cannot resolve independently.

This approach ensures the human remains the leader of the system, using the AI to handle the heavy lifting while retaining final judgment on structural changes. The final step is a bias toward action, emphasizing that the system improves through actual usage and feedback loops rather than over-engineering the initial setup.