Before AI Changes the UI, Tell It a Screen Has Three Readers

A developer on X proposed two AGENTS.md rules for vibecoders. AGENTS.md can be understood as a project instruction file for coding agents; the agent reads it before starting work. The rules are short and very specific.

First: every screen has three readers. The user viewing the interface, the developer changing the interface, and the next agent editing it. All three experiences must be served; otherwise you're tearing one down to patch another.

Second: delete labels that exist only to explain layout. If a screen needs explanatory text to be understood, the screen itself has a problem.

Screenshot of the original post:

The most easily overlooked part here is AX, agent experience. It's not about making the interface look agent-friendly; it's about whether the agent can find the place that needs changing without guessing. In a reply, the author added an image: one screen, three readers. The user sees the rendered result, the developer sees the component file, and the agent must first read AGENTS.md, find SaveButton.tsx, edit it, and then check the screen. When the path is clear, the agent reads fewer files and guesses less about locations.

The author's addition about AX is crucial: the agent should be able to find what to edit without guessing. If changing a button requires reading six files, then even if the result the user sees is perfect, the AX is poor. One component, one obvious point of modification, and what the user sees must not break.

In the comments, one developer says that if the agent doesn't know where things are in the repo, it will guess, and guessing is the most time-consuming. Someone else says that designing only for the people looking at the screen today will miss the people maintaining it tomorrow. Another reply is more direct: AX is often forgotten until an agent breaks a layout that was just fixed.

Some also question whether it's too hard for small teams to handle all three experiences at once. That concern is valid. The two-line rule doesn't add a new process; it moves maintainability one step earlier: let the agent first know how many readers a screen has, then start working.

Source: Khairallah AL-Awady's post https://x.com/eng_khairallah1/status/2117321468841517449