Tone and Brand Voice Constraint Compliance Prompt

The generated copy to evaluate for tone and brand voice compliance

Our new Pro plan gives you unlimited access to advanced analytics and priority support for just $49/month.

Required. Must be non-empty string. Truncate to 4000 tokens if longer to avoid judge attention drift.

The canonical brand voice definition with dimension descriptors and acceptable ranges

Formality: professional but warm (score 3-5). Empathy: acknowledge user context (score 4-5). Assertiveness: confident not pushy (score 2-4).

Required. Must include at least one dimension with numeric range. Missing ranges cause judge to invent thresholds.

2-5 examples of copy that perfectly match the brand voice guidelines

Exemplar 1: We built this feature because you asked for it. Here is how it works. Exemplar 2: Sorry that took longer than expected. Let us make it right.

Required. Minimum 2 exemplars. Fewer than 2 causes calibration drift. Exemplars must be real approved copy, not synthetic.

The specific tone dimensions to score, drawn from the brand voice guidelines

formality, empathy, assertiveness, warmth, clarity

Required. Must be a list of 2-7 dimensions. Each dimension must appear in [BRAND_VOICE_GUIDELINES]. Extra dimensions cause hallucinated criteria.

Per-dimension numeric score ranges that count as compliant

formality: 3-5, empathy: 4-5, assertiveness: 2-4, warmth: 3-5, clarity: 4-5

Required. Must map 1:1 to [DIMENSIONS]. Missing range for any dimension causes that dimension to be skipped or scored arbitrarily.

Number of out-of-range dimensions that trigger an overall non-compliant verdict

2

Required. Integer between 1 and length of [DIMENSIONS]. Set to 1 for strict compliance, higher for tolerance of minor drift.

The expected JSON structure for the judge response

{"overall_compliant": boolean, "overall_score": number, "dimension_scores": {"[dimension]": number}, "violations": [{"dimension": string, "score": number, "expected_range": string, "evidence": string}]}

Required. Must be valid JSON Schema or example structure. Judge prompt should include this inline. Schema mismatch between prompt and harness causes parse failures.

Optional human-annotated scores for exemplars to anchor the judge scale

{"exemplar_1": {"formality": 4, "empathy": 5}, "exemplar_2": {"formality": 3, "empathy": 4}}

Optional. If provided, judge prompt should instruct comparison against these anchors. Missing calibration increases score variance across runs.

number (0.0–1.0)

Must be a float between 0.0 and 1.0 inclusive. Parse check: JSON number. If null or out of range, trigger retry or fallback to human review.

array of objects

Array must contain at least one dimension object. Each object must include dimension_name (string), score (number 0.0–1.0), and justification (string). Schema check: required fields present.

string

Must match one of the brand voice dimensions defined in [BRAND_VOICE_DIMENSIONS] input. Enum check: reject unknown dimension names.

number (0.0–1.0)

Must be a float between 0.0 and 1.0 inclusive. Parse check: JSON number. If null, flag as incomplete evaluation.

string

Must be non-empty and reference specific evidence from [GENERATED_COPY]. Citation check: justification must quote or paraphrase a concrete example from the input text.

array of objects

Array may be empty if no violations found. Each object must include violation_type (string), excerpt (string), expected_behavior (string), and severity (string). Schema check: required fields present when array is non-empty.

string

Must be one of: tone_mismatch, formality_break, empathy_gap, assertiveness_error, register_shift, or other. Enum check: reject unknown types.

string

Must be one of: minor, moderate, major, critical. Enum check: reject unknown severities. If critical severity present, flag for mandatory human review.

string

If present, must reference [BRAND_EXEMPLARS] input and explain how the score aligns with or deviates from exemplar patterns. Null allowed when exemplars are not provided.

Output matches the target formality level specified in [BRAND_VOICE_GUIDE] (e.g., casual, consultative, formal). No more than one register shift per paragraph.

Frequent code-switching between formal and informal registers. Slang or colloquialisms in a formal context, or stiff jargon in a casual context.

LLM-as-judge pairwise comparison against [BRAND_EXEMPLAR] using a 1-5 formality scale. Flag if score deviates by more than 1 point from target.

Output demonstrates the empathy level defined in [EMPATHY_TARGET] (e.g., high for support, low for technical reports). Emotional language is appropriate to the scenario.

Cold or dismissive language in a high-empathy context. Overly sentimental or intrusive language in a low-empathy context. Use of exclamation marks or emoji where prohibited.

Keyword and sentiment analysis for empathy markers. LLM judge rates empathy on a 1-5 scale against [EMPATHY_TARGET]. Fail if score is 2+ points off target.

Output matches the assertiveness profile in [ASSERTIVENESS_PROFILE] (e.g., suggestive, confident, definitive). Hedging is used only where uncertainty is genuine.

Excessive hedging ('might', 'could possibly') in a definitive context. Overconfident declarations ('always', 'never') in a nuanced context. Unqualified claims without evidence.

LLM judge evaluates hedging frequency and confidence markers. Compare ratio of hedging phrases to definitive statements against [ASSERTIVENESS_PROFILE] benchmark. Flag deviation greater than 30%.

Output uses only terminology approved in [TERMINOLOGY_LIST]. Industry terms are used correctly and consistently. No internal slang or competitor terminology.

Use of deprecated or competitor product names. Internal code names or unapproved abbreviations. Incorrect usage of defined terms that changes meaning.

Automated string match against [TERMINOLOGY_LIST] and [DENY_LIST]. LLM judge checks for correct contextual usage of approved terms. Flag any deny-list match or misuse of approved terms.

Output follows the sentence structure patterns defined in [VOICE_EXEMPLARS]: length, complexity, and rhythm match the brand's typical cadence.

Run-on sentences in a brand that uses short, punchy prose. Choppy fragments in a brand that uses flowing, complex sentences. Paragraph length deviates significantly from exemplars.

Compute average sentence length, complexity score, and paragraph length. Compare distribution to [VOICE_EXEMPLARS] corpus. Flag if mean sentence length differs by more than 40% from exemplar mean.

Output uses inclusive language as defined in [INCLUSION_GUIDE]. Gendered terms, ableist language, and culturally specific idioms are avoided unless explicitly approved.

Use of gendered pronouns for unknown individuals. Idioms that assume specific cultural knowledge. Outdated or non-inclusive terminology for groups or conditions.

Automated regex and dictionary scan against [INCLUSION_GUIDE] violation patterns. LLM judge reviews flagged terms for contextual appropriateness. Fail if any unambiguous violation is detected.

Output expresses the top [N] brand personality traits from [BRAND_PERSONALITY] (e.g., witty, authoritative, warm). Traits are evident but not forced or caricatured.

Output is bland and generic with no detectable brand traits. Forced humor that falls flat. Trait expression that contradicts the brand (e.g., sarcasm in a sincere brand).

LLM judge rates presence of each trait in [BRAND_PERSONALITY] on a 1-5 scale. Compare to [TRAIT_TARGETS]. Flag if any required trait scores below 3 or if a prohibited trait is detected.

Output is appropriate for [TARGET_AUDIENCE] in terms of reading level, assumed knowledge, and cultural references. Does not talk down or over the audience's head.

References that assume knowledge the audience lacks. Over-explaining basic concepts to an expert audience. Cultural references irrelevant to the target market or demographic.

Readability score check (e.g., Flesch-Kincaid) against [AUDIENCE_READING_LEVEL]. LLM judge evaluates assumed knowledge level against [AUDIENCE_PROFILE]. Fail if readability is more than 2 grade levels off target.