Skip to content
LinkedInX

What Humans Do When AI Writes the Code: Separating Judgment, Verification, and Design

About This Article

When I started building a site with AI, I did not know where to draw the line on what to delegate, and for a while I tried to have the AI verify every implementation decision. Later, as I grew more comfortable, there were times when I delegated too much and skipped verifications I should have done myself.

This article shares the division between “tasks that can be delegated to AI” and “tasks that humans must verify” that I have worked out through experience.

I leave the pre-delegation decision to the three axes in Criteria for Delegating Tasks to AI. This article focuses on human responsibility for direction, verification, and security decisions during and after implementation.

Tasks That Can Be Delegated to AI

Concrete Implementation (Writing Code)

Tasks with a clear target and change — “change the button color from blue to white,” “add a link to this paragraph,” “add a new item to the navigation” — can be delegated to the AI.

For these tasks, the important question is not “can it be executed” but “what to execute.” That judgment is made by humans. The execution itself is handled by the AI.

Presenting Multiple Options

Having the AI produce options — “list three ways to implement this feature” or “tell me what configuration options are available for this setting” — is an area where AI excels. Information that would take significant time to research independently can be organized quickly.

Investigating the Cause of Errors

When an error occurs, the error message can be passed to the AI and it can explain “what is happening” and “how to address it.” This removes the need to interpret error messages personally.


Tasks That Humans Must Verify

Deciding on Direction

Questions like “does this site need this feature,” “what page structure should I use,” and “is this design direction right” need to be answered by humans. The AI suggests options for “how to implement,” but “what to build” is decided by humans.

Confirming External Service Connection Settings

Configurations linking Vercel to GitHub, connecting to external APIs, and so on can be set up following AI instructions, but the final confirmation of which repository or account is being connected to must be done personally. A mistake in these settings can leave publishing scope or permissions in an unintended state.

Verifying the Display After Publishing

Even if the AI has set everything up correctly, verifying how the live site actually appears in a browser cannot be skipped. There can be layout differences between desktop and smartphone, or navigation links that point to the wrong destination — problems that cannot be caught from configuration files alone.

Judgments like “is this configuration safe to publish” or “where should this information be stored” are ones the AI can offer proposals for, but final confirmation must be handled by humans. Settings involving the management of API keys and passwords in particular require careful review.


The Human Role Is Not the Ability to Write Code

When I started using Vibe Coding, I had the feeling that “I can’t write code, so I’ll end up relying entirely on AI.” But as I worked through actual projects, I came to realize that the ability to write code matters less than “the ability to decide what to build,” “the ability to verify whether what was built matches the intent,” and “the ability to articulate the next instruction.” These factors have a greater impact on the quality of the work.

Before and after the AI’s “implement” step are human steps: “judge,” “verify,” and “instruct.” The quality of those surrounding steps determines the quality of the whole site.

Summary

The division of roles is not fixed — it shifts depending on the nature and context of the work. One useful decision criterion is to ask: “If a mistake happened in this task, could I identify it myself?” The less able I am to verify a task independently, the more important it is to understand the underlying mechanism before delegating it to AI.