claude vs codex: the editor

Last entry covered Codex’s pass on the editor — including a tagline that drifted from the one explicitly defined in the project bible. This one is Claude’s pass on the same surface, plus the call I made about which to keep going with.

Same instruction to both: build whatever you can right now, skip the pieces that depend on me (email notification setup, R2 storage, etc).

Where Codex won:

Codex went straight at the 2D canvas and shipped things I never asked for but immediately wanted:

• Wall continuation. First click sets the start, second click sets the end. Third click extends from the second, and so on — closing a room is just clicking your way around it.
• Grid by default. The working canvas already had a grid. I didn’t ask for one.
• Door alignment. Drop a door on a wall and it sat flush with the wall automatically.

None of these were in the brief. All of them are obviously right. On this front, Codex outclassed Claude — honestly, no contest.

Where Claude stumbled:

Same primitives, clumsier:

• Walls don’t continue. First click is the start, second click is the end, and that’s it. To extend the segment I have to click again to begin a new wall. Codex’s chain-mode is a real ergonomic gap — closed rooms are the whole point of this surface.
• Blank canvas, no grid. Drawing a closed area without a grid is much harder than it should be.
• Doors at strange angles. Drop a door on a wall and it lands rotated, crossing the wall at 90° — a position that makes no sense for a 2D plan.

Not a fatal gap. A prompt or two would close it (probably one, with a tight enough scope). But out of the box, Codex was the better instinctive designer here.

Where Claude pushed back:

The interesting part. When I asked Claude to skip the pieces that depend on me, it didn’t just agree. It pushed back: those skipped pieces are foundations, and pulling them in now makes everything later easier — dropping them is inconsistent with the way the project bible is laid out. It still left the call to me. I followed its advice and stuck with the bible.

Codex took the same skip-the-deps instruction and ran with it. Claude argued. That’s a plus for Claude.

The pick:

Going forward I’m continuing on Claude. I’m not saying Codex is worse — its unprompted instincts on the 2D surface were better than Claude’s, and that’s not nothing. The reason is workflow fit. My development style leans heavily on custom agents and commands tuned to the specific project, and I’m already comfortable with how that lands in Claude.

Not a “Claude is smarter than GPT” claim — that takes more metrics than I can fairly cover here, and most of the smarter-than debate ends up subjective anyway. Both are excellent in their own way.