Don't Fix It With a Rule

I expected the studio's prompts to be carefully hand-authored documents — the crown jewels, written and rewritten by a person. Mostly, a model wrote them. Almost nobody hand-writes a prompt from scratch anymore; you draft it with a model and then you tune it, over and over, as things go wrong in QA. The interesting part isn't the writing. It's the tuning — and there's a trap in it that took the studio a while to see.

The reasonable rule

Here's how the trap springs. A room comes out of the pipeline wrong — too cluttered, say, or too bare. The obvious fix, right there in the moment, is a rule:

text   the reasonable fix                 the durable fix
   ──────────────────                 ───────────────
   ✗ no more than six fixtures        ✓ a room should feel like a place
   ✗ at least two clues per wall      ✓ exploring and solving are one act
   ✗ the finale must be a keypad      ✓ earn one real click of delight
   ( shapes → every room the same )   ( goals → every room its own )

Each rule on the left is reasonable. Each one works — the very next room obeys it. And each one is a small tax the studio doesn't notice it's paying, because the cost isn't in any single room. It's in all of them at once.

Why rules make a model dull

A model with one constraint is still creating. A model with forty is doing something else: it's satisficing. Faced with a long list of hard "must" and "never," it stops hunting for the interesting solution and starts hunting for the safe one — the leanest arrangement that happens to clear every bar at once. Pile up enough perfectly reasonable rules and you get the studio's own hard-won lesson, recorded the day it over-corrected into a rigid checklist and had to climb back out:

A checklist of shapes produces rooms that all look identical because they all satisfy the same checklist.

The rules weren't wrong. There were just too many, and they were the wrong kind.

Principles, not shapes

The fix the studio settled on is one distinction. A rule dictates a shape — how many fixtures, which ending, where the clues go. A principle names a goal — feel like a place, be fair, earn a surprise — and says nothing about the shape that gets you there. Rules close the space a model can move in; principles point across it and leave the route open. The whole design doctrine now fits on one line: optimize the objectives, not the tactics.

Hard rules don't vanish — a few are load-bearing. But they're reserved for the things about quality and solvability, never structure: a puzzle must be fair, an answer must be reachable. Those can be firm without killing anything, precisely because they constrain whether a room works, not what it looks like. Everything that used to be a structural rule became a principle, and the checks that remain mostly warn instead of block: the gate that looks for a room's funnel shape, for one, prints its objection and lets the room through — you have to pass an explicit --strict to make it stop the line. The fairness checks are the exception; those still refuse to build.

So the discipline, when a room fails QA, isn't to add the rule — but it isn't to wave the failure into an abstraction either. It's a specific move, in three parts. Lift the failure into the principle underneath it: the objective it actually missed. Keep the case attached — the principle carries the original bad room with it, as a worked example, so it stays concrete and a reader can see exactly where the line falls. (A principle without its example is a platitude; the reference is what makes it teachable, and what keeps it honest.) And consolidate: reconcile the new principle against all the others, because principles overlap and quietly contradict, and a set nobody keeps coherent rots the same way a checklist does. One principle — carrying its example, reconciled with the rest — replaces a dozen rules you'd have accumulated a bad room at a time, and unlike them it leaves the model somewhere to be surprising.