Juejin's recent long-read from a programmer crossed 10,000 views, and the author's take is blunt: when evaluating AI coding tasks, "can it be done" is the wrong question—"where do the boundaries lie" is the right one.
What this is
The author offers four judgment dimensions: whether the rules are clear, whether results are easy to verify, whether errors are easy to roll back, and whether the cost of failure is controllable. The more "yes" answers, the more suitable the task is for AI to take the first pass; the fewer, the more human oversight is required.
On that basis he sorts tasks into three tiers: directly delegatable (writing tests, organizing logs, drafting documentation); AI analyzes, humans decide (interface refactoring, performance tuning); developer-only (business rules, permissions and money, production environment configuration). The article also includes a copy-paste Prompt template for clarifying the split before each task begins.
Worth noting: this four-question method—rules, verification, rollback, cost—isn't exclusive to programmers. Any scenario involving handing work to AI can use these four questions as a first-pass filter.
Industry view
Reactions in the tech community are split.
Supporters argue this hits the real pain point of enterprise AI rollout—not that models aren't strong enough, but that no one dares take responsibility for the output. Silicon Valley engineering teams have recently offered similar judgments: "AI accountability boundaries" are the core issue for enterprise deployment in 2026, harder to solve than model selection.
The opposing view is worth more attention. A senior architect pointed out in comments that this framework is essentially an extension of "waterfall thinking" (the traditional development model of defining all requirements upfront before starting)—breaking tasks into clear inputs and outputs, then letting AI fill in the blanks. But in reality, much work involves "defining requirements as you go," and AI's advantage lies exactly in that gray zone. Forcing the four questions limits AI's value in the requirements-exploration phase.
Our take: the framework itself is useful, but only if requirements are already clear. If your requirements are still being shaped by AI's output (for example, using AI to explore product form), this framework isn't enough—you need a separate set of "AI exploration phase" collaboration rules.
Impact on regular people
For enterprise IT: Stop asking "can the team use AI to write code" and start asking "which steps do we hand off and which must stay with humans." The four-question framework can drop directly into procurement evaluation tables—more practical than just comparing model parameters.
For individual careers: Even if you don't code, the logic generalizes. Tasks with well-defined rules (organizing meeting notes, drafting reports) are fair game for delegation; anything involving money, contracts, or personnel decisions must have a human call the final shot—AI only lists the options.
For the consumer market: Direct impact is limited. But if AI coding becomes widespread in enterprises, software customization costs will bifurcate—small-batch, rapid-validation tools will get cheaper, while the bar for developing large core systems will rise, because the "judgment" they require becomes scarcer and more expensive.