What a leaked system prompt says about yours

When the Fable system prompt leaked, most writeups read it as a look inside the vendor: the hidden rules, the character instructions, the wording about alignment. You can read it that way. The part that stayed with us was more boring. The document has structure. It has sections, layered rules, and a clear line between types of behavior. Someone worked on it more than once.

Now open the system prompt running in your own production app and be honest about what is in there.

We see the same thing most of the time. A paragraph written during the prototype that nobody has touched since. A few lines on tone, copied from a blog post. Instructions that quietly contradict each other, because every person added a requirement and nobody removed one. No comment that explains why any of it is there. If there is a git history at all, the last commit says “tweak prompt” and the person who wrote it has already left.

A system prompt is not a config string you set once and forget. It is the main control you have over how the model behaves in front of real users. When it is messy, the output is messy in the same places, and you will lose an afternoon blaming the model for a sentence a human wrote back in March.

Treat it like code

The habits you already use on source code mostly just need to point at the prompt.

Put the prompt in git as its own file, so a change comes as a diff and not as a Slack message. Send prompt edits through the same review as everything else. Write a small set of fixtures, an input and the output you expect from it, and run them before anything ships. None of this is hard. We have watched teams with strict review and full CI push a prompt change straight to prod with none of it.

So that is what we would do. Find the prompts that are actually live. Give each one an owner and a changelog. Add a few regression checks, even ugly ones.

The leak is worth copying in one way. Someone could tell you what that prompt does, and would notice the day it stopped doing it. Most teams cannot say that yet, and the prompt is usually doing more work than the code around it.