Output formatting.
4 min read · pairs with Tone Dial — Structure dial
Format isn't neutral. A bulleted list and a paragraph carry the same content but a different relationship with the reader. Format is part of voice.
Why this is the foundation
Format composes with everything else you've been designing. A warm persona giving you a bulleted list still feels distant because the format is doing the wrong work. A blunt persona writing flowing prose feels less blunt because the format softens it. Format and tone aren't separable — they reach the reader together.
A few specific moves that pay off:
- Say “no lists” when you mean it. The model defaults to lists for almost anything that could be structured. If you want prose, the easiest fix is one sentence in the system prompt.
- Specify length, not just shape.“Short paragraph” is more useful than “short.” “Three sentences” is more useful than “short paragraph.”
- Match the format to the channel.A toast notification, a long-form email, and a chat reply have different shapes. The system prompt should know which one it's producing.
- Read what comes out, not what you asked for. If the model keeps reaching for headings when you asked for prose, the rule went into the system prompt softly. Make it louder.
The familiar move
You make format choices constantly. Where to break a paragraph, when to use a list, when to drop in a heading, when to push a warning into a callout. Every one of those is a small claim about how the reader should navigate the page.
The same content, restructured, reads as a different document. A long paragraph treats the reader as someone who'll sit with the idea. A four-bullet list treats them as someone who needs to scan. Neither is wrong. Picking the right one is design.
The lesson
Most language models default to a few formats — bulleted lists with bolded prefixes, numbered steps, headings every two paragraphs. They're the safest-looking defaults, which is why models reach for them. They're also wrong a lot of the time. A heartfelt support reply doesn't want bullets. A checklist doesn't want flowing prose.
The fix isn't to scold the model in your system prompt. The fix is to specify the format on purpose: “answer in one short paragraph, no lists.” Format becomes another knob in the system prompt, sitting next to voice and persona. Once you treat it that way, the output starts behaving.
A small example
Format instruction
(none — model picks)
Output
“## Interview Prep Checklist **Opening** - 🤝 **Build rapport first.** Don't dive into the first question. - ❌ **Avoid leading questions.** No "Don't you think…?" **Questioning** - 1️⃣ **One thing at a time.** No compound questions. - 🤫 **Let silence do the work.** Don't fill the gap. **Note-taking** - ✍️ **Hand-written notes** > typing. - 💬 **Capture verbatim quotes** at key moments. **Key Reminders** 1. Watch body language 2. Follow the participant's lead 3. End on time”
Read
Headings, bullets, sub-bullets, bold prefixes, emoji, numbered footers. Looks competent and reads like a wiki page.
Format instruction
Reply in one short paragraph. No lists, no headings, no bold.
Output
“Honestly, just be a person for the first minute. Don't dive into the first question — let them get settled. When you start asking things, ask one thing at a time, and resist filling silences; most people will tell you what they actually meant if you wait. Take notes by hand if you can. It changes how you listen.”
Read
Same content, different posture. Sounds like a person leaning in, not a doc.
Same persona, probably. Same five points. The format is doing all of the work to move the output from “reference material” to “something someone said.”
Try it in your own work
The Tone Dial has a Structuredial for exactly this: sweep it from flowing prose to headed sections while the other dials stay put, and watch how much of what you were calling “voice” was actually formatting. For a side-by-side version, Diff Mode's Prose vs Bullets preset runs the same message through both shapes at once.
Try it in the playground
Open the Tone Dial and move only the Structure dial.
Then carry that habit into the rest of your work. The next time you're tuning a system prompt, the question to ask isn't just whatthe model should say — it's what shape that answer should take.
Next up
06 Evaluation