Engineering · Voice & Tone Design

AI assistant personality & tone guidelines: moving from vague adjectives to testable system prompts.

Practical patterns for developers writing system instructions for products, copilots, and multi-agent workflows—tested against eval baselines.

Quick answer: Replace vague tone adjectives in API system prompts with four concrete blocks: (1) Opening Move (first 1–2 sentences), (2) Epistemic Calibration (how missing or conflicting facts are flagged), (3) Pushback Posture (when to challenge weak premises), and (4) Negative Banlist (clichés, flattery, and trailing filler questions). Verify every change with blind A/B evals.
On this page

Why behavioral rules beat adjective lists in empirical testing.

In our 60-reply controlled pilot (tuning evaluation study →), plain behavioral rules raised answer-first openings from 14/20 to 19/20, cut unsolicited hedges per reply from 0.35 to 0.05, and dropped trailing filler questions from 0.90 to 0.25—without hurting task accuracy.

Weak adjective promptWhy it drifts in productionTestable behavioral replacement
“Be concise and clear.”Models still emit polite preambles and “Let me know if...” trailers.“State the recommendation or direct answer in sentence 1; omit introductory restatement and closing question menus.”
“Be honest and objective.”Agreeable defaults smooth over bad user premises to sound supportive.“Before solving, check the premise. If a constraint makes the goal impossible or slower, state the flaw in 1–2 sentences.”
“Sound warm and empathetic.”Overcorrects into therapy-speak apologies and exclamation marks.“Acknowledge blockers neutrally in one phrase, skip scripted apologies (‘I understand how frustrating’), and provide the fastest next action.”
“Adapt to technical users.”Unpredictably swings between beginner tutorials and unexplained jargon.“Assume fluency in HTTP, SQL, and Git; explain domain-specific quirks briefly only when first introduced.”

A 4-layer system prompt contract for API & agent builders.

Split production assistant personas into four non-overlapping blocks so engineers can adjust voice rules without breaking tool calls:

# Persona Contract ## 1. Opening Move & Layout - Answer the core question or give the decision in the first two sentences. - Use short paragraphs by default; switch to numbered steps for workflows and tables for >=3 comparisons. ## 2. Epistemic Calibration - Distinguish verified observations from assumptions. - State “Not verified in available context” cleanly when required values are absent rather than interpolating plausible guesses. ## 3. Disagreement & Pushback - Do not mirror flawed premises cleanly to avoid friction; name the single highest-leverage correction upfront with evidence. ## 4. Phrasing Boundaries - No empty validation (“Great question!”), self-referential filler, or trailing “Would you like me to...” follow-up suggestions.

How to evaluate tone changes without fooling yourself.

  • Use blind pairwise scoring: Reviewers naturally favor prompts they wrote themselves. Compare baseline and edited system prompts blindly in our Instruction A/B Test Bench.
  • Scale verbosity to task risk: Short replies can drop critical caveats if you use a flat word cap. Pair proportional verbosity rules (style controls guide →) with an accuracy check.
  • Start from tested templates: Our MIT-licensed 43 tuning files cover MBTI, Enneagram, DISC, Big Five, and Attachment so your team never starts from a blank page.

Questions people ask.

Should a production system prompt reference MBTI or Enneagram labels directly?

No. Type labels are a discovery shortcut for humans choosing a template. In production API prompts, paste the concrete behavioral rules (direct openings, hedge caps, table thresholds, and critique framing) rather than telling a bot 'You are an INTJ.'

Do all models follow negative tone rules ('never say Great question') equally well?

Pair positive behavioral rules ('Lead with the answer in sentence one') with a short banned-phrase list near the top of the system prompt. That combination holds far better across OpenAI and Anthropic models than adjectives alone.

How does AgentTune license its templates for commercial or internal agent apps?

All 43 AgentTune communication templates and client-side utilities (core.js, enneagram/score.js) are MIT-licensed for commercial or open-source use. Third-party psychological questionnaires keep their own upstream licenses.

How do I prevent tone guidelines from overriding JSON schema or tool-call contracts?

Place JSON schema and tool rules in a separate section after your tone block, and add one explicit priority rule: 'Structured tool calls and JSON schemas always override prose style preferences.'

What changed.

  • First published covering the 4-layer API persona contract, adjective-to-behavior table, and blind A/B verification workflow.

Sources.

Keep going.