Roles, Personas, and System Prompts
Leverage roles and system instructions to shape expertise, tone, and boundaries across single and multi-agent setups.
Content
Voice, Style, and Tone
Versions:
Watch & Learn
AI-discovered learning video
Sign in to watch the learning video for this topic.
Voice, Style, and Tone — Make Your AI Sound Human(ish)
You already picked the right role. You calibrated expertise. You wrote crisp, unambiguous instructions. Now ask the model to say it like a person who actually knows how to talk.
We’ve covered "Selecting Effective Roles" and "Calibrating Expertise Levels" — great! That set the who and the how smart. This chapter is the how they sound. Think of it as choosing the persona's wardrobe, accent, and preferred expletives (PG-13 version). Build on your clear instructions: attach a voice that fits the role and acceptance criteria.
Tiny definitions that save careers
- Voice — the consistent personality behind the words. (e.g., "concise domain expert", "friendly coach", "sarcastic techie").
- Style — the mechanical choices: sentence length, jargon use, punctuation, formatting habits (bullets, code blocks, analogies).
- Tone — emotional flavor from line to line: empathetic, urgent, excited, skeptical.
Why separate them? Because you can have the voice of a professor, the style of a Twitter thread, and the tone of someone comforting you after a failed lab report. Each layer is tweakable and composable.
Why this matters (a short, slightly dramatic truth)
If role + expertise = the brain,
then voice + style + tone = the face it shows the world.
Mismatched voice breaks trust. A "Senior Legal Counsel" persona writing like a frat bro is confusing. Worse: unclear tone can make acceptance criteria fail. You asked for "concise summary" and got a poetic novel — acceptance: NOT met.
How to design voice/style/tone (step-by-step)
- Start with the Role & Expertise Level (we already did this). Ask: "Who is speaking?" and "How deep are they?"
- Define the audience: novice, peer, executive, client. Voice changes by audience.
- Specify style constraints using your previous checklist technique: scope, constraints, acceptance criteria. Add voice/style-specific constraints.
- Add tone cues (emotional palette + intensity).
- Provide examples (good/bad) and a tiny rubric for evaluation.
Example constraint you already use for instructions → extend it:
- Scope: 150–250 words summary aimed at product managers
- Constraints: no jargon, use numbered steps, include one analogy
- Acceptance criteria: <1 jargon term, 3 actionable bullets, readability score: easy
- Voice/style addition: voice: "calm PM mentor"; style: short sentences, active voice; tone: encouraging, slightly urgent.
Voice-to-prompt mapping (cheat-sheet)
| Persona / Role | Voice Traits (short) | Style Examples | System prompt snippet |
|---|---|---|---|
| Junior Tutor (novice audience) | patient, encouraging | short sentences, analogies, no jargon | "You are a patient tutor who explains step-by-step like teaching a 10th grader." |
| Senior Researcher (peer audience) | precise, terse, evidence-focused | formal, citations, dense language | "You are a senior researcher; prioritize clarity and cite sources." |
| Product Manager Coach | pragmatic, outcome-focused | bullets, prioritization, business framing | "You are a PM coach: give prioritized 3-step plans and trade-offs." |
| UX Copywriter | friendly, microcopy-savvy | microcopy examples, A/B variants | "You are a UX writer; propose 3 UI strings, each 2–6 words." |
Tone Palette — pick one (and be brave)
- Formal / Neutral: professional, low emotion. Use for legal, academic.
- Conversational / Warm: approachable, uses contractions, metaphors. Use for onboarding.
- Urgent / Direct: short sentences, command verbs, time-based language. Use for alerts.
- Empathetic / Reassuring: acknowledges feelings, soft phrasing. Use for customer support.
- Playful / Witty: jokes, rhetorical questions, casual slang. Use sparingly.
Quick examples (tone applied to same instruction):
- Formal: "Please complete the following report by 5 PM."
- Conversational: "Could you wrap up the report by 5? That would be awesome."
- Urgent: "Report due 5 PM — prioritize now."
- Empathetic: "I know it's a lot, but can you finish the report by 5? We're here to help."
- Playful: "Let’s hit 5 PM on that report and celebrate with virtual cake."
System Prompt Templates (plug & play)
Short template you can drop into your system prompt:
You are {role} with {expertise_level}. Speak in a {voice} voice. Use {style_constraints}. Tone should be {tone}. Follow acceptance criteria: {acceptance_criteria}. Provide examples and justify trade-offs briefly.
Filled example:
You are a Senior Data Scientist with deep expertise. Speak in a concise, evidence-first voice. Use bullet points, include a 2-sentence summary, one code snippet, and one citation. Tone: objective and slightly urgent. Acceptance: ≤250 words, 1 code sample that runs, 2 actionable recommendations.
Quick rubric to evaluate voice/style/tone (useful for QA)
Score 0–2 on each:
- Consistency (Is voice consistent across the reply?)
- Appropriateness (Does it match role + audience?)
- Clarity (Does style help readability?)
- Emotional fit (Does tone convey the intended feeling?)
Accept if total ≥7/8. If <7, iterate: tighten system prompt, add examples, or simplify constraints.
Common mistakes + one-line fixes
- Model is too verbose —> add: "Limit to X words; use bullets; avoid background unless asked."
- Model uses heavy jargon for novices —> add: "No jargon. If necessary, define a term in one sentence."
- Tone drifts mid-answer —> add: two short examples (good/bad) and enforce "maintain tone throughout."
- Voice doesn't match expertise —> tie voice to expertise explicitly: "As a mid-level engineer, explain with practical code snippets and avoid academic prose."
Micro-recipes (5 one-liners you can paste)
- "You are a friendly product coach. Summarize the trade-offs of X in 5 bullets, plain English, no jargon."
- "You are an impatient bug triager. Report root cause in 3 lines, include reproduction steps and a recommended priority."
- "You are an empathetic customer support agent. Apologize, explain the fix in simple terms, and offer next steps."
- "You are a witty marketing copywriter. Give 6 tagline variants, each ≤6 words, playful tone."
- "You are a clinical editor. Rewrite this paragraph to be formal, concise, and neutral."
Closing (the part that makes you feel smarter)
Voice, style, and tone are the emotional safety rails on top of your technical instructions. You already know how to pick the right role and expertise level and write clear acceptance criteria — now make sure the way the model says things matches that brain you built. Test with the rubric, iterate with examples, and remember: a role that knows a lot but speaks like a motivational poster probably won't get the job done.
Quote to remember:
"Clarity is the caterpillar; voice is the butterfly. You need both to fly."
Go forth and give your models a personality that actually helps people.
Comments (0)
Please sign in to leave a comment.
No comments yet. Be the first to comment!