Xinzhiyuan · WeChatMay 1, 2026, 13:29 UTC+8

Claude Code's Real Story: 98.4% of What Works Is Engineering, Not AI

撕开Claude Code真相：让它好用的98.4%，是工程不是AI

VILA-Lab analyzed 512,000 lines of Claude Code v2.1.88 and found 1.6% tied to AI decision logic. The other 98.4% is deterministic infrastructure: permissions, context, tool routing, and error recovery. The key shift is harness design, not longer prompts.

Why it matters: Strong HKR: the Claude Code teardown has a sharp counter-narrative and concrete 512k LOC plus 1.6%/98.4% split. It is not an official Anthropic release and lacks full reproduction details, so it stays in the 78–84 band.

Read the original ↗Export Markdown

Take

Claude Code’s exposed number is the knife here: 1.6% of 512,000 lines is described as AI decision logic, while the rest is permissions, context management, tool routing, and error recovery. The WeChat body is blocked by verification, so the methodology, file classification, and dependency boundaries are not verifiable. Treat 98.4% as a directional claim, not a clean audit result.

I buy the direction. The gap between coding agents over the last year has come less from mystical prompts and more from whether the product can run shell commands, edit a repo, recover from failure, and keep context under control. OpenAI Codex, Cursor, and Claude Code are all fighting in that wrapper layer. A stronger model gets you in the room; a safer harness decides whether an engineering team lets it touch production code.

Source: Xinzhiyuan · WeChat · mp.weixin.qq.com
Published: May 1, 2026, 13:29 UTC+8 · May 1, 2026