Prompts as design.
6 min read · pairs with Diff Mode
A prompt is a design variable, not a magic spell. Once you see it that way, most of the “prompt engineering” folklore starts to feel familiar — because you already do this work, just with different inputs.
Why this is the foundation
Behavior design is mostly small loops. Write a version, watch the model behave, notice the gap between what you wanted and what you got, change one thing. The size of the loop matters: the smaller and faster, the more design you do.
Diff Mode shrinks the loop. Two configs, one user message, both outputs side-by-side. You stop debating Version A vs Version B in your head and start seeing them.
The familiar move
When you write a tooltip, you don't hand the engineer a paragraph and hope. You write a brand-voice document, decide on a tone (warm but not cute, direct but not curt), then craft the copy to match. The brand-voice doc is the persistent style; the tooltip is the call site.
The system prompt is the brand-voice doc for the model. The user message is the call site. Same architecture — different surface.
The lesson, stated plainly
Every word in a system prompt is a design decision. “Be concise” and “Be brief” are different style choices. “You are a research assistant” and “You are an experienced research assistant who works with first-time interviewers” produce different outputs.
When you treat the prompt as a design variable, the work becomes recognizably yours: write a version, run it, look at the output, notice what you don't like, change one thing, run it again. That's iteration, not incantation.
A small example
System prompt
You are a research assistant.
Output
“Welcome! How can I help you with your research today?”
System prompt
You are a research assistant who's been doing this for fifteen years. You've watched a lot of interviews fail because the interviewer asked a leading question in the first minute. Be warm, but cut straight to what matters.
Output
“Hi — quick check before we get going: do you have one or two specific behaviors you're hoping to see today? If not, that's our first move.”
Same task. Same model. Different prompt, different behavior. You didn't need to know anything about transformers or tokens to see why Version B is more useful — you read the two outputs and one felt right.
Try it in the playground
Open Diff Mode and change one variable.
What to take into the playground
- Start with a prompt you've actually written before — a microcopy guideline, a voice doc, a brief. Don't invent.
- Change one variable per run. Two variables and you won't know which one moved the output.
- Save the run as a Diff Log. The point isn't the answer; it's the trail of decisions you made.
Next module
02Voice & tone