What this is
Juejin author "fly" updates Chapter 6 of his "AI Native Full-Stack Engineer" course, moving into hands-on execution. Students connect their own Codex (OpenAI's coding Agent), then build and deliver a complete group-buying project — organizer launches a deal, buyers place orders, order fulfillment.
The course introduces an engineering concept: Harness — students' first assignment is to build this "engineering foundation," which includes code checks, submission records, read-back materials (documents that map actual behavior against requirements line by line), and the verification relationships with the instructor. The course hammers one slogan: "The stronger the model, the more Harness you need."
In other words, AI writing code is no longer novel. Putting AI-generated code back into the PRD (Product Requirements Document) and checking each line item against it — matching version numbers, matching the verifier — that's what this course actually teaches.
Industry view
Supporters: AI coding tools have evolved to the Cursor, Claude Code, Codex generation. The bottleneck has shifted from "can't write code" to "trustworthy delivery." Turning Harness into a process and a product is exactly the necessary path for enterprises landing AI coding in 2025 — the course direction aligns with industry observation.
Skeptics: This is a single author's experimental course. Whether the method applies to large teams, compliance scenarios, or cross-department collaboration remains unverified. AI coding tools churn every six months; the Harness designed around Codex today will likely need rewriting tomorrow. A more practical concern: the course emphasizes "each student gets an independent repository with independent verification," which carries high replication cost and offers limited takeaways for enterprise training.
Impact on regular people
For enterprise IT: If your company is evaluating AI coding tools, a more useful question than "can it write code?" is "how do we verify and audit?" This course walks through that mechanism with reasonable specificity.
For individual careers: People who "can direct AI to write code" are no longer scarce. What's starting to pay is "people who can judge whether AI wrote it right and leave evidence behind." The earlier you build this capability line, the stronger the moat.
For the consumer market: AI coding courses will keep flooding in. The litmus test is simple — courses that teach "how to use tools" are everywhere; courses that teach "how to verify, how to own delivery" are the ones worth paying for.