The first system prompt I ever wrote for a client had four adjectives before the verb.
"You are an experienced, insightful, strategic, and creative thinker."It read like a job description written by someone who’d interviewed the role and liked the candidate. It also did almost nothing. I didn’t understand that yet. I understood it a few hundred instruction sets later. The prompt sounded perfect. It was also, functionally, wallpaper.
I was listening to Taylor Swift talk about songwriting, and she may have accidentally described how to write a great system prompt or agent instruction set.
“Details should not be strictly decorative.”
The more I work with AI agents, system prompts, and increasingly complex instruction sets, the better I’ve gotten at telling the difference between a detail that’s doing something and one that’s just fancy dress.
My Substack - Some Assembly Required - has always been where I work things out in public. It’s where the ideas are longest, sharpest, and least edited for palatability. If you want to stay connected to where the thinking is going, not just the occasional post that survives the algorithm, please subscribe.
Casting Notes, Not Instructions
“You’ve got that James Dean daydream look in your eye.” - Style. From the record “1989”, Side A, Track 3
Think about what a specific detail does inside a good song, instead of a mediocre one. “You’ve got that James Dean daydream look in your eye” is not decoration. It’s retrieval. Two words and you’ve loaded an entire system: rebellious, cinematic, doomed, a little dangerous, gone too soon. The specificity isn’t there to sound nice. It’s there to do work the writer would otherwise need a paragraph to do, and possibly not as well.
Now reread the system prompt I opened with. Experienced. Insightful. Strategic. Creative. Every one of those words describes a person I’d want in the room. None of them tells the model how to behave once it’s there.
Delete “insightful” — does the output change? Delete “experienced” — what computation shifts? I’ve run this test on dozens of client instruction sets since, and the answer is almost always: nothing happens. You’ve written a casting note. You haven’t designed a behavior.
Consider the kind of system prompt we’ve all written:
You are an experienced strategic thinker. You are insightful and creative. You communicate clearly and concisely. You challenge conventional thinking. You are an expert in marketing, culture and consumer behavior.Sounds pretty good. Sounds specific. Sounds like someone we’d probably want in the meeting. But here’s the rub. We’ve written a lovely character description, but again, we have not designed behavior.
Decorative Specificity vs. Consequential Specificity
There’s another Swift lyric I love:
“The dopamine races through his brain on a six-lane Texas highway.” - I Can Fix Him (No Really I Can). From the record “The Tortured Poets Department.” Side C, Track 3
A highway is a detail. A six-lane Texas highway is doing considerably more. It gives you scale, speed, heat, geography, a particular kind of American restlessness. Take out “six-lane” and “Texas” and you haven’t just made the line shorter. You’ve made the man less specific.
Decorative specificity makes a prompt sound like it knows what it wants. Consequential specificity changes what the model does when it gets there.
“You are an experienced strategic thinker” is decorative. It describes the kind of person I would like to have in the room. But it doesn’t give that person anything specific to do.
“Before recommending a strategy, identify the assumption that must be true for it to work” is different. Now something has to happen before the recommendation appears.
“You are decisive and opinionated” sounds good too. “Never present three equivalent recommendations. Evaluate the alternatives and recommend one” is better. The first is a flattering character reference. The second changes everything that follows.
Same with objectivity. “You are objective and evidence-based” is a nice thing to say about someone. “When evidence conflicts with the user’s hypothesis, privilege the evidence” gives the agent a rule for an actual moment of conflict.
One describes the agent. The other designs its behavior. That’s the difference.
Every Detail Should Change What Happens Next
“Midnight / You come and pick me up, no headlights.” - Style. From the record “1989”, Side A, Track 3
“No headlights” changes everything. Remove it and someone simply picks her up at midnight. Keep it, and suddenly there is secrecy, risk, illicitness. Two words alter the inferred situation.
Which brings us back to Taylor Swift.
“Details should not be strictly decorative.”
What if we applied that literally to every instruction we put into a system prompt? Every sentence should earn its place by changing what happens next.
It should affect a decision. Alter a priority. Introduce a constraint. Trigger a behavior. Change how ambiguity is resolved. Determine which evidence wins. Influence when a tool is used. Tell the agent when to stop. Tell it when not to act. Otherwise, what exactly is the sentence doing there?
This doesn’t mean descriptive language is useless. Context matters. Roles matter. Tone matters. Sometimes telling an AI who it is supposed to be creates a useful prior for everything that follows.
But there is a difference between establishing context and accumulating adjectives. And we’re doing an awful lot of accumulating adjectives.
The Taylor Swift Test
Here’s the test I actually use now, because “does this sound like something a smart person would write” is a terrible filter, and “would this change the next output” is a good one.
Pick a sentence in your instruction set. Delete it. Then test the agent in the situation the sentence is supposed to govern. If its behavior doesn’t change in any meaningful way, it was probably never an instruction. It was costume.
When I think about system or agent design, I usually look out for three things:
Decorative: Sounds desirable but does not reliably alter behavior. “You are insightful.”
Consequential: Changes what happens at a specific decision point. “When evidence conflicts with the user’s hypothesis, privilege the evidence.”
Priming: Does not operate as a hard rule, but establishes a useful default across the work: role, audience, source hierarchy, or a genuine stance towards the task.
When Words Become Actions
With a chatbot, a decorative instruction might not matter very much. Maybe the answer is slightly more generic. Maybe the tone wanders. Maybe you get three recommendations when you really wanted one.
But agents change the stakes. Because an agent doesn’t merely generate language. It can make decisions. It can select tools. It can retrieve information. It can decide what information matters. It can execute sequences. It can determine whether something is complete.
Increasingly, it can act. Which means the system prompt sits upstream of an entire decision tree. A seemingly innocuous sentence can change not merely what the AI says, but what it does next. This makes instruction design less like writing a creative brief and increasingly like designing a behavioral system. And behavioral systems don’t need decoration. They need consequence.
I’ve spent a chunk of this year arguing that orchestration, not prompting, is becoming the real technical skill in this work — that the craft has moved from writing a good sentence to designing a system that makes good decisions on its own.
Maybe Better Prompts Aren’t Longer
There’s another lesson from songwriting here that I think matters. Great songwriting doesn’t necessarily contain more detail. It contains the right detail. The detail that carries disproportionate meaning. The detail that allows the listener to infer everything else. The detail that makes an entire emotional world suddenly visible. Which makes me wonder whether we’re heading in the wrong direction with system prompts.
As agents become more sophisticated, the temptation is to make their instructions increasingly enormous. More rules. More context. More adjectives. More edge cases. More paragraphs explaining what the agent should be. Until eventually we’ve written something resembling an employee handbook for a machine.
Maybe that’s not where great agent design goes. Maybe the real craft will be learning to write fewer sentences with greater consequence. Instructions that don’t merely describe desirable behavior. Instructions that produce it. Details that change decisions. Details that create causality. Details that make the next action inevitable.
Taylor Swift wasn’t talking about AI. She was talking about songwriting. But “details should not be strictly decorative” may be one of the better principles I’ve encountered for designing intelligent systems.
So the next time you’re staring at a 3,000-word system prompt, try this. Pick a sentence you like. Delete it.
If nothing changes, you didn’t just find a redundant line. You found out how much of the document was never instructional.
Maybe it was just a really nice costume.
