An open-source project called "skills" gained 23k Stars in three weeks, revealing a clear judgment: the bottleneck in AI coding has shifted from generation speed to engineering discipline.
What this is
Matt Pocock, a renowned educator in the TypeScript field, open-sourced his engineering judgment process for coding with AI, forming 21 Agent (AI programs capable of autonomous task execution) skills. He proposed a systematic solution to the currently rampant vibe coding (coding by humans and AI based on vague feelings, advancing in a shared illusion).
This method doesn't teach prompt tricks, but diagnoses typical AI coding failure modes and provides targeted remedies: against AI understanding deviations, use structured follow-up questions to clear ambiguities; against AI verbosity and cognitive load, establish a shared language (like a unified project glossary) to reduce communication costs; against AI writing non-runnable code, forcefully introduce TDD (Test-Driven Development: writing test cases before implementation code) to rebuild the feedback loop. The core idea: map and enforce traditional software engineering discipline into AI collaboration.
Industry view
We note that senior engineers widely agree with Pocock's diagnosis. AI can mess up a codebase in three days, while human teams need three months to fix it. When coding speed is no longer the bottleneck, controllability must be the top priority. Teaching AI engineering principles rather than fancy instructions is the necessary path from toy projects to production environments.
But it is worth our concern that there are opposing voices: this solution demands extremely high engineering literacy from developers themselves. If users don't understand Domain-Driven Design (a software design method focusing on core business logic) or testing specifications, these skills become cumbersome process shackles that actually slow down development. Furthermore, overly constraining the AI's execution path might also stifle emergent, unexpectedly effective solutions from the model in certain scenarios.
Impact on regular people
For enterprise IT: When evaluating the ROI of AI coding tools, post-maintenance costs must be factored in. Fast AI generation does not equal overall cost savings; they must be paired with engineering specification reviews.
For individual careers: The shortcut for junior programmers to quickly produce code via AI is failing. Engineering capabilities—understanding architecture design and testing specifications—are the true career moats.
For the consumer market: In the short term, software bugs caused by AI-rushed development may increase. In the long run, only AI constrained by engineering discipline can deliver stable and reliable products.