Reverse-engineer a prompt from an output you liked

I have an output I liked. **Work backwards to the prompt that would reproduce it.**

1. **Describe what the output actually is** — ***its structure, not its content.* **Sections, order, length per part, what each section does**
2. **Find the rules it follows**, ***including ones it follows silently*** — **what it never does, how it handles uncertainty, where it stops**
3. ***Separate what came from the prompt from what came from the input.*** **This is the hard part** — *a detail that looks like a prompt rule may just be a property of that one input*
4. **What made it good** — ***name it.* **"It was clear" is not a specification**

**Then write the prompt:**
- **Task, constraints, output shape, and what to do when information is missing**
- ***Do not embed this output as an example*** unless the shape is the whole point — **and if you do, strip its specific values**
- **Mark what I would need to confirm** by running it on a different input

**Then:** ***what will not reproduce.* **Some of what I liked was this input, or luck. Say which parts those probably are.**