Simon Willison — Django framework co-founder and long-time independent developer tracking AI tools — published a short post this week. His core conclusion: the key to using Coding Agents (AI assistants that can autonomously write code) isn't line-by-line code review, but rather "confident instruction and confident verification." We see this as the first clearly articulated "soft skills manual" for the Agent era.
What this is
Simon Willison has an outstanding reputation in overseas developer circles — his blog has run for twelve years, and he carries the credibility of being a Django co-founder. The single core sentence from his post this week: the key skill for getting Coding Agents to actually deliver is confident instruction plus confident verification, not line-by-line review.
He further argues that eyeballing code line by line has never been the most effective way to validate software changes. This is counter-intuitive — over the past year, "AI-written code must be manually reviewed" has become near industry political correctness. Simon's message: don't spend your effort in the wrong place.
Industry view
Plenty of support. Karpathy and others have been repeatedly raising the "AI engineer" concept lately — essentially the same point Simon is making: future developers' core competitive edge is judgment, not code-typing speed.
But the objections also hold. A senior architect noted in public comments: "Simon himself is a top-tier developer; he can verify because he actually understands the code. For newcomers, line-by-line review is a lifeline." Others point out that in strong-compliance scenarios such as healthcare and finance, line-by-line audit isn't optional — it's a legal requirement. We believe both sides are right; the key depends on scenario and team maturity. Simon's experience can't be directly transplanted to beginners.
Impact on regular people
For enterprise IT: job postings may shift from "proficient in language X" to "able to decompose fuzzy requirements into AI-comprehensible steps." Verification capability will be worth more than coding capability.
For individual careers: even non-programmers can use AI to write scripts and generate reports. The barrier isn't technical — it's articulating the problem clearly and verifying the result correctly.
For the consumer market: this may give rise to "AI verification" tools and training courses. This category is still a blank slate; first movers capture the dividend.