What Happened
A Chinese developer community post on Juejin is gaining traction for naming a real problem: Vibe Coding, the practice of iteratively prompting AI without any written specification. The author proposes Spec-Driven Development (SDD) as the structured alternative. SDD requires three artifacts before any code is written: a proposal.md explaining the why, a specs/ directory containing versioned acceptance criteria, and a tasks.md checklist that locks scope. One internal team cited in the article measured 2.3 unintended side effects per feature change under Vibe Coding. Under SDD, AI is restricted to executing only what appears in tasks.md — nothing outside the spec is touched.
Why It Matters
Indie developers and small teams using AI coding assistants face compounding debt from unscoped changes. A typical failure mode: asking an AI to "optimize login" results in the entire auth module being rewritten, including password recovery flows nobody asked to change. SDD addresses this with a four-step workflow: draft proposal → review and align → AI executes against locked spec → archive updated master spec. The key constraint is that AI cannot modify anything not explicitly listed in tasks.md. This makes regression behavior predictable and gives teams a single source of truth stored in version control.
- Specs live in
specs/as Markdown files with checkable acceptance criteria - Each change gets its own folder:
changes/feature-name/containing proposal, tasks, and spec diffs - AI role shifts from creative collaborator to contract executor
- Rollback becomes a spec revert, not a code archaeology exercise
Asia-Pacific Angle
Chinese and Southeast Asian developers building SaaS products for global markets often work in small teams where one developer handles both product decisions and implementation. SDD is particularly useful here because the spec files double as product documentation for international users and potential investors — written in English, they communicate feature behavior without requiring live demos. Teams using Cursor or Claude for code generation can paste the entire specs/ directory as context, which significantly improves output quality compared to conversational prompting. This also helps when outsourcing to freelancers across Vietnam, Indonesia, or the Philippines: the spec file is the contract, reducing miscommunication from language and timezone gaps.
Action Item This Week
Pick one feature you are currently building with AI assistance. Before your next prompt session, create a changes/your-feature/ folder with three files: a proposal.md (two sentences on why this change is needed), a specs/feature.spec.md with three to five acceptance criteria written as checkboxes, and a tasks.md listing every file the AI is allowed to modify. Paste all three files as context at the start of your AI session and instruct the model to only touch files listed in tasks.md.