The real secret behind Claude Code is probably not the model

Claude Code's source code leaked, revealing the secret behind its performance. Rather than simply relying on the model's capability, the core lies in the Software Harness that uses the model efficiently.

The key technical features that make Claude Code more powerful than the web UI are as follows.

• Real-time repository context building: Loads Git branch, recent commits, CLAUDE.md, and more to instantly grasp the working environment.
• Aggressive prompt cache reuse: Optimizes cost and processing time by separating and caching static content from dynamic content.
• Advanced tool utilization: Uses dedicated Grep, Glob, and LSP (Language Server Protocol) tools instead of simple Bash commands, improving the accuracy of file search and code structure analysis.
• Minimizing context bloat: Solves context limitation issues through deduplication of file reads, saving large results to disk with previews provided, and automatic summarization.
• Structured session memory: Systematically records work status, tasks, errors, and fixes in markdown format, managed similarly to how a human would code.
• Fork and subagent utilization: Designed to parallelize tasks and perform background analysis or summarization without affecting the main loop.

In conclusion, Claude Code's excellence stems not from prompt engineering or the model's performance alone, but from sophisticated performance optimization and efficient context management capability.