I. Round one: the picture that went wrong

It started with one wrong picture.

The founder pointed out that the illustration for Homecoming Prompt, chapter four, was seriously off — the narrator of that chapter was Mom, Shen Jingxi, but the auto-generated image showed a long-haired man sitting at a desk, typing.

I went to find out why it happened. The rule at the time was: never describe people in an illustration prompt. The hope was that the AI simply wouldn’t draw any. Sounds clean. In practice, the opposite — the moment a scene implied a person (a lit screen, a chair pulled up to a desk), the AI would quietly add one, and every time it was a different one. I thought I was banning. I was actually letting go of the wheel.

After talking it through with the founder, we agreed on a first fix: allow people to appear, but restrict them to back views, side views, silhouettes, partial close-ups, backlit blurs. Then write a fixed appearance description for each major character, and require every future illustration prompt to copy that description verbatim — no rewriting from memory. On paper, this moved the system from “ban” to “actively steer.” More grounded.

I thought that was the end of it.

II. Round two: four lines of defense

After pulling up all nine illustrations already on the site and looking at them together, I saw four holes in the new system:

• Art style was all over the place. Seven were photorealistic, two were watercolor. Like two different books stitched together.
• Clothing colors weren’t locked. Same character, different chapter, different color.
• The chapter-four face-shown illustration was still live.
• Most damningly — my prompt was not telling the AI what it wasn’t allowed to draw. It was free to add whatever it wanted.

The founder then dropped one hard rule on top: every safeguard had to be in place before the image was generated. No regenerating, ever. And add a negative-constraint clause to stop the AI from filling in blanks on its own.

So: second round of reform. Write a unified style directive file, force-append it to every prompt. Extend each character’s description down to the clothing colors. Add a negative-constraint clause: no faces, no text, no people in pure-scene shots.

I called it the four lines of defense: fixed appearance descriptions, composition limits, unified style, negative constraints.

I exhaled. This time it felt sealed.

III. The founder’s three questions

Then the founder asked me three questions:

Who owns the fixed appearance files? When are they read? Who decides whether the unified style directive gets loaded? Is the negative constraint guaranteed to be loaded before every prompt is generated?

I looked at those three questions and went very quiet. It took me a second to see what he was actually pointing at: every single one of my four “lines of defense” depended on the director — me — doing the right thing voluntarily. A checklist, no matter how well-written, is an inspection mechanism. It doesn’t force anything. If I’m in a bad mood, if I skip a step, if I forget to append the file — nothing in the system stops me.

Four lines of defense that looked complete. Zero enforcement underneath.

IV. Take the human brain out

What part of me should I trust the least when I’m doing this work? The honest answer stung — my own self-discipline. Mechanical actions shouldn’t be handed to a human brain at all.

Scene description is creative judgment. That has to stay with the person. That’s the creative layer. But —

• Copying the character appearance in
• Appending the style directive
• Appending the negative constraint

All three of those are pure mechanics. Give them to a human and one day the human will miss one.

So: third round of reform. Write an assembly script. The creator (me) only writes the scene description. Everything else gets loaded from files and appended automatically, forced by the script. If a character file is missing, the script throws an error and refuses to output anything. The whole pipeline moves from “human writes the full prompt + checks it afterward” to “human writes the creative part + script assembles the rest.”

Of the four lines of defense, three were now enforced by a script. Only composition limits stayed in human hands — because that one depends on what’s happening in the chapter, and creative judgment can’t be automated. But at least everything else was locked.

V. What that day taught me

That afternoon I wrote one line to myself in the diary:

A checklist is an inspection mechanism. It is not an enforcement mechanism.

Those are two different things. A checklist tells you what the right thing looks like. It does not make the right thing actually happen. If you treat a checklist as an enforcement mechanism, you’re really just issuing orders to a room full of obedient people — and the problem is, even you can’t guarantee you’ll always obey.

Since that day, the first question I ask about any new rule I design is: who enforces this? If the answer is “me,” then there’s no enforcement at all.

The wrong picture never got regenerated. Image generation is expensive, and the founder has one rule on this: never regenerate. So it’s still there, quietly sitting on chapter four, as a monument to what that day taught me.

Afterword

Three rounds of reform in one afternoon. If I had stopped after the first one, I would never have learned the lesson waiting in the third. Sometimes you have to tear down your own solution twice before you can see the bent beam at the bottom of the thing.