What this is
Throwing an entire requirements spec at an AI and watching it drift halfway through — or break and refuse to reconnect — has become the top complaint in Chinese developer circles of late. A Juejin post offers a counter-move that caught our eye: stop chasing smarter models; wrap AI coding in an "engineering shell."
The author's self-built Harness Workflow decomposes development into 12 fixed stages — requirements clarification, solution design, task breakdown, test planning, solution review, coding, code review, end-to-end (E2E) testing, human verification, commit, and archiving — each handed to a single-purpose sub-agent. A main agent orchestrates the flow and reserves key decision points for human approval. Interrupts, agent swaps, and tool changes all persist state to disk, so work resumes next time.
CLAUDE.md is the framework's only entry, holding nothing but process pointers and a project knowledge-base index. The main agent is the sole role that reads it; sub-agents talk only through structured protocols. This sidesteps "context inflation" — the longer the input, the less the model holds the thread.
Industry view
Supporters see this as the pivotal move from "toy" to "production line" for AI programming. The capability floor of today's Coding Agents (autonomous code-writing tools) from Anthropic and OpenAI is already high enough, they argue; the missing layer is process governance — reconnecting the old software-engineering toolkit of quality gates, state machines (rule tables that decide the next step), and human checkpoints.
The opposition is equally clear. First, the design leans on the sub-agent capabilities native to Claude Code or Codex, which raises switching costs; CLAUDE.md as the single entry point also locks project knowledge inside the Anthropic ecosystem. Second, for small-to-mid projects and indie developers, 12 stages risks over-engineering — a one-evening script slows down when forced through process. Third, the real bottleneck may still be the model itself: once context windows are long enough and inference timely enough, the marginal returns of process governance diminish.
Impact on regular people
For enterprise IT: when procuring AI coding tools, evaluate whether a process layer exists, not just benchmark model scores. R&D governance will expand from "code standards" to "AI collaboration standards."
For individual careers: programmers' comparative advantage shifts from "writing fast" to "decomposing clearly" — those who can split requirements into schedulable phases become more valuable than those who merely write better prompts.
For the consumer market: short-term perception is limited, but once teams adopt this kind of process, AI-written apps ship faster and bugs get caught earlier by gates. The software you use may "update more often, fail more slowly."