SDD Stops AI Coding from Eating Itself: A Chrome Extension Walkthrough
AI code generation is cheap, but clear, executable intent is scarce. SDD gives teams a repeatable way to produce that intent before the AI touches code, cutting the rework cycle that makes vibe coding economically irrational for anything beyond toys.
Vibe coding — telling an AI to build something and iterating in chat — skips the mental blueprint and jumps straight to code, which collapses into rework and hallucination by week two. SDD restores that first creation with three structured documents: a proposal that defines what to build and what not to build, a design that locks in architecture and tech choices, and a task list that breaks work into single, verifiable units. The approach was applied to a real Chrome extension that extracts English articles via Mozilla Readability, translates them through an OpenAI-compatible API (model-swappable between DeepSeek and Qwen), and renders the result as WeChat-optimized Markdown with typewriter streaming. Every AI session runs one task only, starts with a clean context, and every change is immediately version-controlled so hallucinations can be rolled back. When a new requirement arrived — switching from a popup to a full-height side panel — the docs were updated first, then the code followed, keeping spec and implementation in lockstep under git.
The core economic argument is inverted from what most developers assume: as code generation gets cheaper, the bottleneck shifts upstream to specification quality, not downstream to coding speed.
SDD treats documentation not as a side artifact but as the primary engineering artifact — 'documentation is code' — which changes where senior effort should be spent in an AI-assisted workflow.
The single-task-per-session rule is a practical countermeasure to the context-window decay that silently degrades AI output quality across long conversations.
Immediate git versioning of AI output turns hallucinations from catastrophic surprises into cheap, reversible mistakes, which changes the risk calculus of letting an agent generate large diffs.
The side-panel iteration demonstrates that SDD's real payoff is not the first build but the Nth change: when docs and code stay coupled, every future modification has a verified starting point.