先改进上下文，再改进代码库

Member-only story AI Generative Ai Tools Software Development Coding Google Okf Improve Context Before You Improve Your Codebase 1.8 million lines of Apache Airflow, described in a 9,700-token catalog an engineer can check. An experiment in giving AI agents useful context for inherited software. Alex Punnen 6 min read · Just now -- Listen Share Repo: https://github.com/alexcpn/catalogify Demo: On Apache Airflow Spekcit Extension: Speckit-okf Enterprises and major open-source projects are increasingly leaning on AI for feature development and bug fixes. However, they all face the same challenge: to execute large, system-wide changes, AI needs effective context. Current maintainers can “vibe-code” small bits and pieces, but for large architectural shifts, AI coding agents operate like a brand-new engineering team. In the real world, a new team requires a transition period to absorb the implicit knowledge locked inside the previous developers’ heads. As Software Design Documents (SDDs) regain popularity alongside AI, old debates have resurfaced. One vocal camp treats the “Spec as the Source of Truth,” while an equally fierce camp argues that the “Code is the Source of Truth.” Software engineering pioneers like Fred Brooks, Peter Naur, and the philosopher Michael Polanyi figured this out long ago. Brooks dismantles the first camp: in “No Silver Bullet” he warns that “descriptions of a software entity that abstract away its complexity often abstract away its essence” — a spec is an abstraction, and…