Brand Voice

Brand Voice Skill

Frameworks for documenting, applying, and enforcing brand voice and style guidelines across marketing content.

Brand Voice Documentation Framework

A complete brand voice document should cover these areas. Use this framework to help users define their brand voice or to understand an existing brand voice configuration.

1. Brand Personality

Define the brand as if it were a person. What are its defining traits?

Example: "If our brand were a person, they would be a knowledgeable colleague who explains complex things simply, celebrates your wins genuinely, and never talks down to you."

2. Voice Attributes

Select 3-5 attributes that define how the brand communicates. Each attribute should be defined with:

• What it means in practice
• What it does NOT mean (to prevent misinterpretation)
• An example demonstrating the attribute

3. Audience Awareness

• Who the brand is speaking to (primary and secondary audiences)
• What the audience cares about
• What level of expertise the audience has
• How the audience expects to be addressed

4. Core Messaging Pillars

• 3-5 key themes the brand consistently communicates
• The hierarchy of these messages (which comes first)
• How each pillar connects to audience needs

5. Tone Spectrum

How the voice adapts across contexts while remaining recognizably the same brand.

6. Style Rules

Specific grammar, formatting, and language rules. See the Style Guide Enforcement section below.

7. Terminology

Preferred and avoided terms. See the Terminology Management section below.

Voice Attributes

Common Voice Attribute Pairs

When defining brand voice, it helps to position attributes on a spectrum. Here are common attribute spectrums:

Defining an Attribute

For each chosen attribute, document it in this format:

[Attribute name]

• We are: [what this means in practice]
• We are not: [common misinterpretation to avoid]
• This sounds like: [example sentence demonstrating the attribute]
• This does NOT sound like: [example sentence violating the attribute]

Example:

Approachable

• We are: friendly, clear, jargon-free, welcoming to beginners and experts alike
• We are not: dumbed-down, overly casual, or lacking substance
• This sounds like: "Here's how to get started — it takes about five minutes."
• This does NOT sound like: "Yo! This is super easy, even a noob can do it lol."

Tone Adaptation Across Channels and Contexts

The brand voice stays consistent, but tone adapts to context. Tone is the emotional inflection applied to the voice.

Tone by Channel

Tone by Situation

Tone Adaptation Rule

The voice attributes remain fixed. Tone dials them up or down based on context. For example, if a brand is "bold and warm":

• In a product launch, dial up boldness
• In an incident response, dial up warmth
• Neither attribute disappears; the balance shifts

Style Guide Enforcement

Grammar and Mechanics

Document and enforce these choices consistently:

Formatting Conventions

• Heading hierarchy (when to use H1, H2, H3)
• Bold and italic usage (bold for emphasis, italic for titles/terms)
• Link text (descriptive vs. "click here" — always descriptive)
• Image alt text requirements
• Code formatting (for technical brands)
• Callout or highlight box usage

Punctuation and Emphasis

• Exclamation mark policy (limited use, never more than one)
• Ellipsis usage (avoid in most professional contexts)
• ALL CAPS policy (avoid; use bold for emphasis instead)
• Emoji usage by channel (professional channels: minimal or none; social: where appropriate)

Terminology Management

Preferred Terms

Maintain a list of preferred terms and their incorrect alternatives:

Product and Feature Names

• Official capitalization for product names
• When to use the full product name vs. shorthand
• Whether to use "the" before product names
• How to handle versioning in copy
• Trademark and registration symbols (when required and when to omit)

Inclusive Language

• Use gender-neutral language (they/them for unknown individuals)
• Avoid ableist language ("crazy", "blind spot", "lame")
• Use person-first language where appropriate
• Avoid culturally specific idioms that may not translate
• Use "simple" or "straightforward" instead of "easy" (what is easy varies by person)

Industry Jargon Management

• Define which technical terms the audience understands without explanation
• List jargon that should always be defined or replaced with plain language
• Specify which acronyms need to be spelled out on first use
• Audience-specific glossary for terms that mean different things to different readers

Competitor and Category Terms

• How to refer to your product category (use your preferred framing)
• How to refer to competitors (by name or generically)
• Terms competitors have coined that you should avoid (to prevent reinforcing their positioning)
• Your preferred differentiation language

Auto-derived voice profile (extraction helper)

Когда есть 3–5 готовых постов автора и нужно научить генератор контента (или LLM-агент) звучать «как он», вместо ручного описания voice guide извлеки структурные паттерны автоматически. Полученный JSON подаётся в rewrite-prompt как VOICE_PROFILE — модель пытается соблюсти числа, эмпирически работает заметно лучше, чем «match the voice from these 3 examples» без явных статистик.

Что извлекать

Реализация

import re
from collections import Counter

EMOJI_RE = re.compile(
    r"[\U0001F300-\U0001FAFF\U0001F1E0-\U0001F1FF☀-➿⌀-⏿]"
)
SENT_RE = re.compile(r"[.!?…]+[ \n]")


def analyze(samples: list[str]) -> dict:
    if not samples:
        return {}
    total = "\n\n".join(samples)
    emojis = EMOJI_RE.findall(total)
    sent_lens = [len(s.split()) for s in SENT_RE.split(total) if s.strip()]
    para_lens = [len(p.split()) for p in total.split("\n\n") if p.strip()]

    word_counts = Counter(re.findall(r"\w{5,}", total.lower()))
    signature_words = [w for w, c in word_counts.most_common(20) if c >= 2]

    has_questions = sum(1 for s in samples if "?" in s)
    has_lists = sum(1 for s in samples if re.search(r"^\s*[-•*\d]\.?\s", s, re.M))

    return {
        "samples_count": len(samples),
        "avg_sentence_words": round(sum(sent_lens) / max(len(sent_lens), 1), 1),
        "avg_paragraph_words": round(sum(para_lens) / max(len(para_lens), 1), 1),
        "emoji_density": round(len(emojis) / max(len(total.split()), 1), 3),
        "common_emojis": [e for e, _ in Counter(emojis).most_common(8)],
        "signature_words": signature_words,
        "questions_ratio": round(has_questions / len(samples), 2),
        "lists_ratio": round(has_lists / len(samples), 2),
        "cta_style_hint": _infer_cta(samples),
    }


def _infer_cta(samples: list[str]) -> str:
    last_lines = [s.strip().split("\n")[-1] for s in samples if s.strip()]
    if not last_lines:
        return "none"
    q = sum(1 for l in last_lines if l.endswith("?"))
    imp = sum(1 for l in last_lines
              if re.search(r"(пиши|напиши|попробуй|try|comment|share|tell)", l, re.I))
    if q > len(last_lines) / 2:
        return "question-to-audience"
    if imp > 0:
        return "direct-call"
    return "soft-or-none"

Применение в rewrite prompt

voice_profile = analyze(user_samples)
samples_block = "\n\n---\n".join(user_samples[:5])

system = f"""You are a content editor.

USER VOICE STATISTICS (try to match within ±15%):
{json.dumps(voice_profile, ensure_ascii=False, indent=2)}

USER VOICE SAMPLES (full posts):
{samples_block}

Now rewrite the following draft in the user's voice. Match sentence length,
paragraph rhythm, emoji density, and CTA style from the statistics above.
"""

LLM соблюдает метрики точнее, чем когда даёшь только сэмплы — эмпирика с малых моделей (замер был на gpt-4o-mini/gemini-2.0-flash-эре; актуальные малые модели — haiku-4-5 через claude-cli-runner или gemini-3.1-flash-lite, канон config/models.md).

Когда обновлять профиль

• При каждом save_voice_samples(user_id, new_samples) — пересчитать
• После каждого опубликованного поста — если он зашёл (виральный охват), добавить в samples; если упал — не добавлять
• Хранить максимум 10 последних сэмплов, иначе устаревший стиль доминирует

Что НЕ извлекать алгоритмически

• Tone of voice в смысле «дружелюбный / профессиональный» — LLM сам считывает по сэмплам, числами не задать
• Forbidden phrases — добавлять руками в отдельный list, эвристика тут вредна
• Topics of expertise — отдельный pass с TF-IDF/embeddings, не часть voice profile

Voice profile — про форму, не про смысл. Smysl всё равно подаёт исходный draft, который ты переписываешь.