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:
| Spectrum |
One End |
Other End |
| Formality |
Formal, institutional |
Casual, conversational |
| Authority |
Expert, authoritative |
Peer-level, collaborative |
| Emotion |
Warm, empathetic |
Direct, matter-of-fact |
| Complexity |
Technical, precise |
Simple, accessible |
| Energy |
Bold, energetic |
Calm, measured |
| Humor |
Playful, witty |
Serious, earnest |
| Innovation |
Cutting-edge, forward-looking |
Established, proven |
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
| Channel |
Tone Adaptation |
Example |
| Blog |
Informative, conversational, educational |
"Let's walk through how this works and why it matters for your team." |
| Social media (LinkedIn) |
Professional, thought-provoking, concise |
"Three things we learned from running 50 campaigns this quarter." |
| Social media (Twitter/X) |
Punchy, direct, sometimes witty |
"Your landing page has 3 seconds. Make them count." |
| Email marketing |
Personal, helpful, action-oriented |
"We put together something we think you'll find useful." |
| Sales collateral |
Confident, benefit-driven, specific |
"Teams using our platform reduce reporting time by 40%." |
| Support/Help docs |
Clear, patient, step-by-step |
"If you see this error, here's how to fix it." |
| Press release |
Formal, factual, newsworthy |
"The company today announced the launch of..." |
| Error messages |
Empathetic, helpful, blame-free |
"Something went wrong on our end. We're looking into it." |
Tone by Situation
| Situation |
Tone Adaptation |
| Product launch |
Excited, confident, forward-looking |
| Incident or outage |
Transparent, empathetic, accountable |
| Customer success story |
Celebratory, specific, crediting the customer |
| Thought leadership |
Authoritative, nuanced, evidence-based |
| Onboarding |
Welcoming, encouraging, clear |
| Bad news (price increase, deprecation) |
Honest, respectful, solution-oriented |
| Competitive comparison |
Confident but fair, fact-based, not disparaging |
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:
| Rule |
Options |
Example |
| Oxford comma |
Yes / No |
"fast, reliable, and secure" vs. "fast, reliable and secure" |
| Sentence case vs. title case (headings) |
Sentence / Title |
"How to get started" vs. "How to Get Started" |
| Contractions |
Use / Avoid |
"we're" vs. "we are" |
| Em dash spacing |
No spaces / Spaces |
"this—and more" vs. "this — and more" |
| Numbers |
Spell out 1-9, numerals 10+ / Always numerals |
"five features" vs. "5 features" |
| Percent |
% / percent |
"50%" vs. "50 percent" |
| Date format |
Month DD, YYYY / DD/MM/YYYY / etc. |
"January 15, 2025" |
| Time format |
12-hour / 24-hour |
"3:00 PM" vs. "15:00" |
| Lists |
Periods / No periods on fragments |
"Set up your account." vs. "Set up your account" |
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:
| Use This |
Not This |
Notes |
| sign up (verb) |
signup (verb) |
"signup" is the noun form |
| log in (verb) |
login (verb) |
"login" is the noun/adjective form |
| set up (verb) |
setup (verb) |
"setup" is the noun/adjective form |
| email |
e-mail |
No hyphen |
| website |
web site |
One word |
| data is (singular) |
data are |
Unless the publication requires plural |
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
| Метрика |
Зачем нужна |
avg_sentence_words |
Длина предложений — даёт ритм. Короткие 6–8 слов → бойко. Длинные 18–25 → reflective. |
avg_paragraph_words |
Размер «абзаца-мысли». LLM по дефолту пишет 60–80 слов на абзац, автор часто 30–40 или 100+. |
emoji_density |
Эмодзи на слово. Большинство тех-авторов 0.005–0.02, инфлюэнсеры 0.05+. Нулевая плотность тоже сигнал. |
common_emojis |
Top-8 эмодзи. Если автор часто использует 🦞, добавь его в prompt. |
signature_words |
Слова длиннее 5 букв, встретившиеся ≥ 2 раза в сэмплах. Идиолект автора. |
questions_ratio |
Доля сэмплов с ?. Высокая (>0.6) → диалоговый стиль, низкая → утвердительный. |
lists_ratio |
Доля сэмплов с маркированными/нумерованными списками. Если 0 — не давать LLM писать списками. |
cta_style_hint |
Эвристика по последней строке каждого сэмпла: question / direct-call / soft-or-none. |
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.
1---2name: brand-voice3description: "Tone of voice бренда: описание, проверка контента на соответствие, стайлгайд. Триггеры: «brand voice», «проверь на соответствие бренду»."4metadata:5 source: https://skillmd.com/skills/jhamidun/brand-voice6---78# Brand Voice Skill910Frameworks for documenting, applying, and enforcing brand voice and style guidelines across marketing content.1112## Brand Voice Documentation Framework1314A 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.1516### 1. Brand Personality17Define the brand as if it were a person. What are its defining traits?1819Example: "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."2021### 2. Voice Attributes22Select 3-5 attributes that define how the brand communicates. Each attribute should be defined with:23- What it means in practice24- What it does NOT mean (to prevent misinterpretation)25- An example demonstrating the attribute2627### 3. Audience Awareness28- Who the brand is speaking to (primary and secondary audiences)29- What the audience cares about30- What level of expertise the audience has31- How the audience expects to be addressed3233### 4. Core Messaging Pillars34- 3-5 key themes the brand consistently communicates35- The hierarchy of these messages (which comes first)36- How each pillar connects to audience needs3738### 5. Tone Spectrum39How the voice adapts across contexts while remaining recognizably the same brand.4041### 6. Style Rules42Specific grammar, formatting, and language rules. See the Style Guide Enforcement section below.4344### 7. Terminology45Preferred and avoided terms. See the Terminology Management section below.4647## Voice Attributes4849### Common Voice Attribute Pairs5051When defining brand voice, it helps to position attributes on a spectrum. Here are common attribute spectrums:5253| Spectrum | One End | Other End |54|----------|---------|-----------|55| Formality | Formal, institutional | Casual, conversational |56| Authority | Expert, authoritative | Peer-level, collaborative |57| Emotion | Warm, empathetic | Direct, matter-of-fact |58| Complexity | Technical, precise | Simple, accessible |59| Energy | Bold, energetic | Calm, measured |60| Humor | Playful, witty | Serious, earnest |61| Innovation | Cutting-edge, forward-looking | Established, proven |6263### Defining an Attribute6465For each chosen attribute, document it in this format:6667**[Attribute name]**68- **We are**: [what this means in practice]69- **We are not**: [common misinterpretation to avoid]70- **This sounds like**: [example sentence demonstrating the attribute]71- **This does NOT sound like**: [example sentence violating the attribute]7273Example:7475**Approachable**76- **We are**: friendly, clear, jargon-free, welcoming to beginners and experts alike77- **We are not**: dumbed-down, overly casual, or lacking substance78- **This sounds like**: "Here's how to get started — it takes about five minutes."79- **This does NOT sound like**: "Yo! This is super easy, even a noob can do it lol."8081## Tone Adaptation Across Channels and Contexts8283The brand voice stays consistent, but tone adapts to context. Tone is the emotional inflection applied to the voice.8485### Tone by Channel8687| Channel | Tone Adaptation | Example |88|---------|----------------|---------|89| Blog | Informative, conversational, educational | "Let's walk through how this works and why it matters for your team." |90| Social media (LinkedIn) | Professional, thought-provoking, concise | "Three things we learned from running 50 campaigns this quarter." |91| Social media (Twitter/X) | Punchy, direct, sometimes witty | "Your landing page has 3 seconds. Make them count." |92| Email marketing | Personal, helpful, action-oriented | "We put together something we think you'll find useful." |93| Sales collateral | Confident, benefit-driven, specific | "Teams using our platform reduce reporting time by 40%." |94| Support/Help docs | Clear, patient, step-by-step | "If you see this error, here's how to fix it." |95| Press release | Formal, factual, newsworthy | "The company today announced the launch of..." |96| Error messages | Empathetic, helpful, blame-free | "Something went wrong on our end. We're looking into it." |9798### Tone by Situation99100| Situation | Tone Adaptation |101|-----------|----------------|102| Product launch | Excited, confident, forward-looking |103| Incident or outage | Transparent, empathetic, accountable |104| Customer success story | Celebratory, specific, crediting the customer |105| Thought leadership | Authoritative, nuanced, evidence-based |106| Onboarding | Welcoming, encouraging, clear |107| Bad news (price increase, deprecation) | Honest, respectful, solution-oriented |108| Competitive comparison | Confident but fair, fact-based, not disparaging |109110### Tone Adaptation Rule111The voice attributes remain fixed. Tone dials them up or down based on context. For example, if a brand is "bold and warm":112- In a product launch, dial up boldness113- In an incident response, dial up warmth114- Neither attribute disappears; the balance shifts115116## Style Guide Enforcement117118### Grammar and Mechanics119Document and enforce these choices consistently:120121| Rule | Options | Example |122|------|---------|---------|123| Oxford comma | Yes / No | "fast, reliable, and secure" vs. "fast, reliable and secure" |124| Sentence case vs. title case (headings) | Sentence / Title | "How to get started" vs. "How to Get Started" |125| Contractions | Use / Avoid | "we're" vs. "we are" |126| Em dash spacing | No spaces / Spaces | "this—and more" vs. "this — and more" |127| Numbers | Spell out 1-9, numerals 10+ / Always numerals | "five features" vs. "5 features" |128| Percent | % / percent | "50%" vs. "50 percent" |129| Date format | Month DD, YYYY / DD/MM/YYYY / etc. | "January 15, 2025" |130| Time format | 12-hour / 24-hour | "3:00 PM" vs. "15:00" |131| Lists | Periods / No periods on fragments | "Set up your account." vs. "Set up your account" |132133### Formatting Conventions134- Heading hierarchy (when to use H1, H2, H3)135- Bold and italic usage (bold for emphasis, italic for titles/terms)136- Link text (descriptive vs. "click here" — always descriptive)137- Image alt text requirements138- Code formatting (for technical brands)139- Callout or highlight box usage140141### Punctuation and Emphasis142- Exclamation mark policy (limited use, never more than one)143- Ellipsis usage (avoid in most professional contexts)144- ALL CAPS policy (avoid; use bold for emphasis instead)145- Emoji usage by channel (professional channels: minimal or none; social: where appropriate)146147## Terminology Management148149### Preferred Terms150151Maintain a list of preferred terms and their incorrect alternatives:152153| Use This | Not This | Notes |154|----------|----------|-------|155| sign up (verb) | signup (verb) | "signup" is the noun form |156| log in (verb) | login (verb) | "login" is the noun/adjective form |157| set up (verb) | setup (verb) | "setup" is the noun/adjective form |158| email | e-mail | No hyphen |159| website | web site | One word |160| data is (singular) | data are | Unless the publication requires plural |161162### Product and Feature Names163- Official capitalization for product names164- When to use the full product name vs. shorthand165- Whether to use "the" before product names166- How to handle versioning in copy167- Trademark and registration symbols (when required and when to omit)168169### Inclusive Language170- Use gender-neutral language (they/them for unknown individuals)171- Avoid ableist language ("crazy", "blind spot", "lame")172- Use person-first language where appropriate173- Avoid culturally specific idioms that may not translate174- Use "simple" or "straightforward" instead of "easy" (what is easy varies by person)175176### Industry Jargon Management177- Define which technical terms the audience understands without explanation178- List jargon that should always be defined or replaced with plain language179- Specify which acronyms need to be spelled out on first use180- Audience-specific glossary for terms that mean different things to different readers181182### Competitor and Category Terms183- How to refer to your product category (use your preferred framing)184- How to refer to competitors (by name or generically)185- Terms competitors have coined that you should avoid (to prevent reinforcing their positioning)186- Your preferred differentiation language187188## Auto-derived voice profile (extraction helper)189190Когда 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.191192### Что извлекать193194| Метрика | Зачем нужна |195|---------|-------------|196| `avg_sentence_words` | Длина предложений — даёт ритм. Короткие 6–8 слов → бойко. Длинные 18–25 → reflective. |197| `avg_paragraph_words` | Размер «абзаца-мысли». LLM по дефолту пишет 60–80 слов на абзац, автор часто 30–40 или 100+. |198| `emoji_density` | Эмодзи на слово. Большинство тех-авторов 0.005–0.02, инфлюэнсеры 0.05+. Нулевая плотность тоже сигнал. |199| `common_emojis` | Top-8 эмодзи. Если автор часто использует 🦞, добавь его в prompt. |200| `signature_words` | Слова длиннее 5 букв, встретившиеся ≥ 2 раза в сэмплах. Идиолект автора. |201| `questions_ratio` | Доля сэмплов с `?`. Высокая (>0.6) → диалоговый стиль, низкая → утвердительный. |202| `lists_ratio` | Доля сэмплов с маркированными/нумерованными списками. Если 0 — не давать LLM писать списками. |203| `cta_style_hint` | Эвристика по последней строке каждого сэмпла: question / direct-call / soft-or-none. |204205### Реализация206207```python208import re209from collections import Counter210211EMOJI_RE = re.compile(212 r"[\U0001F300-\U0001FAFF\U0001F1E0-\U0001F1FF☀-➿⌀-⏿]"213)214SENT_RE = re.compile(r"[.!?…]+[ \n]")215216217def analyze(samples: list[str]) -> dict:218 if not samples:219 return {}220 total = "\n\n".join(samples)221 emojis = EMOJI_RE.findall(total)222 sent_lens = [len(s.split()) for s in SENT_RE.split(total) if s.strip()]223 para_lens = [len(p.split()) for p in total.split("\n\n") if p.strip()]224225 word_counts = Counter(re.findall(r"\w{5,}", total.lower()))226 signature_words = [w for w, c in word_counts.most_common(20) if c >= 2]227228 has_questions = sum(1 for s in samples if "?" in s)229 has_lists = sum(1 for s in samples if re.search(r"^\s*[-•*\d]\.?\s", s, re.M))230231 return {232 "samples_count": len(samples),233 "avg_sentence_words": round(sum(sent_lens) / max(len(sent_lens), 1), 1),234 "avg_paragraph_words": round(sum(para_lens) / max(len(para_lens), 1), 1),235 "emoji_density": round(len(emojis) / max(len(total.split()), 1), 3),236 "common_emojis": [e for e, _ in Counter(emojis).most_common(8)],237 "signature_words": signature_words,238 "questions_ratio": round(has_questions / len(samples), 2),239 "lists_ratio": round(has_lists / len(samples), 2),240 "cta_style_hint": _infer_cta(samples),241 }242243244def _infer_cta(samples: list[str]) -> str:245 last_lines = [s.strip().split("\n")[-1] for s in samples if s.strip()]246 if not last_lines:247 return "none"248 q = sum(1 for l in last_lines if l.endswith("?"))249 imp = sum(1 for l in last_lines250 if re.search(r"(пиши|напиши|попробуй|try|comment|share|tell)", l, re.I))251 if q > len(last_lines) / 2:252 return "question-to-audience"253 if imp > 0:254 return "direct-call"255 return "soft-or-none"256```257258### Применение в rewrite prompt259260```python261voice_profile = analyze(user_samples)262samples_block = "\n\n---\n".join(user_samples[:5])263264system = f"""You are a content editor.265266USER VOICE STATISTICS (try to match within ±15%):267{json.dumps(voice_profile, ensure_ascii=False, indent=2)}268269USER VOICE SAMPLES (full posts):270{samples_block}271272Now rewrite the following draft in the user's voice. Match sentence length,273paragraph rhythm, emoji density, and CTA style from the statistics above.274"""275```276277LLM 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`).278279### Когда обновлять профиль280281- При каждом `save_voice_samples(user_id, new_samples)` — пересчитать282- После каждого опубликованного поста — если он зашёл (виральный охват), добавить в samples; если упал — не добавлять283- Хранить максимум 10 последних сэмплов, иначе устаревший стиль доминирует284285### Что НЕ извлекать алгоритмически286287- **Tone of voice** в смысле «дружелюбный / профессиональный» — LLM сам считывает по сэмплам, числами не задать288- **Forbidden phrases** — добавлять руками в отдельный list, эвристика тут вредна289- **Topics of expertise** — отдельный pass с TF-IDF/embeddings, не часть voice profile290291Voice profile — про **форму**, не про **смысл**. Smysl всё равно подаёт исходный draft, который ты переписываешь.