humanize-ru
Removes signs of machine generation from Russian-language AI text, preserving its original register and communicative purpose. Does not turn everything into conversational style — scientific text remains scientific, business text remains business, and conversational remains conve
Install
npx skills add https://github.com/v0lka/skills/tree/main/writing/humanize-ru
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install v0lka-skills@llmmart
git clone https://github.com/v0lka/skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole v0lka/skills collection as a plugin from our marketplace. Git is the plain clone.
README
Humanize RU
A standalone Agent Skill for text transformation and post-editing of AI-generated content. It removes signs of machine generation from Russian-language AI text while preserving its original register and communicative purpose — scientific text stays scientific, business text stays business, conversational stays conversational. Only the "AI smell" is removed.
When to use
| Situation | Use this skill |
|---|---|
| Text from an LLM (ChatGPT, Claude, Gemini, …) needs to be published under human authorship | ✅ |
| A client or instructor checks the text with an AI detector | ✅ |
| Machine output sounds dry, formulaic, or unnatural for its genre | ✅ |
| AI output needs to be adapted to a specific tone of voice in Russian | ✅ |
How it works
The skill follows a four-step procedure built around one core principle: not "make the text conversational", but "remove AI markers, keeping the register appropriate for the context."
Step 0 Step 1 Step 2 Step 3
Determine ──► Diagnose ──► Transform ──► Verify
the register AI markers (respecting (overcorrection
the register) test)
Step 0 — Determine the register
The appropriate transformations depend entirely on the functional style:
| Register | Genre examples |
|---|---|
| Разговорный | Telegram post, personal letter, comment |
| Публицистический | Blog, media article, column, longread |
| Научный | Scientific article, dissertation, textbook, report |
| Официально-деловой | Order, contract, business letter, instruction |
| Художественный | Story, novel, essay, literary sketch |
Register is law. Never lower the register without an explicit reason — a scientific text "humanized" to conversational level is a ruined scientific text.
Steps 1–3 — Diagnose, transform, verify
- Diagnose — explicitly list every AI marker found. Transformation does not begin until the markers are enumerated.
- Transform — apply techniques selectively, based on which markers were found and which register was determined. Transformations must not change the register.
- Verify — apply the mandatory overcorrection test: "Could this text have been written by a human in this genre?" If a scientific article sounds like a TikTok post, roll back.
Key principles
- Do not harm meaning. Style is secondary to content.
- Register is law. Do not lower the register.
- Do not overdo it. Conversational elements work on contrast: ≤1 particle per paragraph in publicist text, 0 in scientific and business.
- Verify facts. Better to leave a general formulation than invent a number.
- Organic unevenness > mechanical evenness.
Bundled resources
| File | Purpose |
|---|---|
SKILL.md |
Skill definition, the four-step procedure, and universal rules |
references/ai-markers.md |
Full catalog of AI markers with examples and explanations |
references/transformations.md |
Transformation techniques with dosage tables per register |
references/checklist.md |
Verification checklist with register-dependent criteria |
Compatibility
The skill is agent-agnostic. It describes what to do, not which specific tools to call, so it works with any AI agent that can read and write text. No external services are required.
The skill follows the Agent Skills specification v1.
License
MIT
Skill manifest
Humanize RU — removing AI markers from Russian text
Remove signs of machine generation from the text, preserving its original register and communicative purpose.
Core principle: not "make the text conversational", but "remove AI markers, keeping the register appropriate for the context."
When to use
- Text from ChatGPT, Claude, Gemini, or another LLM needs to be published under human authorship.
- A client or instructor checks the text with an AI detector.
- Machine text sounds dry, formulaic, unnatural for its genre.
- AI output needs to be adapted to a specific tone of voice.
Procedure
Step 0 — Determine the text register
FIRST, determine the functional style (register) of the source text. The appropriate transformations — and the ones that would ruin the text — depend on this.
Five registers and their key features:
| Register | Lexical features | Syntax | "I" allowed? | Genre examples |
|---|---|---|---|---|
| Разговорный | Conversational vocabulary, particles, contractions, ellipsis | Short sentences, parcellation, inversions | Yes, freely | Пост в Telegram, личное письмо, комментарий |
| Публицистический | Common vocabulary + expressive elements, rhetorical devices | Varied: alternating short and long, questions, exclamations | Yes, author's position | Блог, статья в СМИ, колонка, лонгрид |
| Научный | Terms, abstract vocabulary, verbal nouns | Complex syntax, passive voice, clear logical structure | Only "мы" (author's) in methodology | Научная статья, диссертация, учебник, аналитический отчёт |
| Официально-деловой | Bureaucratic formulas, business clichés, standardized phrasing | Cumbersome, clipped, unemotional, strict word order | No | Приказ, договор, деловое письмо, инструкция, отчёт |
| Художественный | Figurative vocabulary, metaphors, individual-authorial words | Free, subordinated to artistic purpose | Depends on narrator | Рассказ, роман, эссе, художественный очерк |
How to determine the register (check in order, first match is the answer):
- Formal marker? Приказ, договор, инструкция, официальное письмо → Официально-деловой.
- Scientific vocabulary? Terms, citations, methodology, abstract concepts as subjects → Научный.
- Literary devices? Plot, characters, explicit imagery, metaphors as foundation → Художественный.
- Personal voice + expression? "Я", rhetorical questions, value judgments, calls to action — without rigid scientific structure → Публицистический.
- Short, informal, with particles? "Ну", "вот", "короче", ellipsis, emoji → Разговорный.
If unsure — default to Публицистический. This is the neutral register, from which you can shift in any direction at the user's request.
NEVER lower the register without an explicit reason. A scientific text "humanized" down to a conversational level is a ruined scientific text.
Step 1 — Diagnosis: find AI markers
Find and explicitly list all signs of machine origin in the text. The full marker catalog with examples and explanations is in references/ai-markers.md. Load this file before starting diagnosis.
Brief overview of marker categories:
- Lexical clichés — formulaic phrases: "в современном мире", "следует отметить", "играет важную роль", "представляет собой" and others.
- Structural patterns — uniform paragraph openings, "rule of three", identical sentence length, template contrasts "не X, а Y".
- Syntax — monotonous word order, genitive chaining, overuse of "который", excessive passive voice.
- Style — facelessness, emotional neutrality, encyclopedic style, absence of conversational elements, inappropriate metaphors.
- English calques — "является" instead of zero copula, "делать" as a universal verb, mandatory subject, "иметь" instead of specific verbs.
- Content signs — lack of specifics, logical loops, absence of context, hallucinations.
CRITICAL: do not proceed to transformation without listing the found markers explicitly. Output the result in the following format:
### Diagnostic Result
**Lexical clichés:** [list of found markers or "none found"]
**Structural patterns:** [list: "rule of three", "не X, а Y", uniform paragraph openings, ...]
**Syntax:** [list: monotonous sentence length, genitive chaining, ...]
**Style:** [list: flat tone, explanatory intonation, absence of personality, ...]
**English calques:** [list or "none found"]
**Content signs:** [list: lack of specifics, logical loops, ...]
**Total markers:** N (3+ = text almost certainly machine-generated)
Step 2 — Transformation: removing AI markers while respecting the register
Apply techniques SELECTIVELY: which markers were found + which register was determined.
Key rule: transformation must NOT change the register. Scientific text remains scientific, conversational remains conversational. Remove only the unnaturalness of AI generation, not the genre traits of the register.
The full catalog of techniques with dosage tables by register is in references/transformations.md. Load this file before transformation. Below are the critically important rules that must not be violated:
Universal rules (all registers):
- "Осуществить" → "сделать" (conv.), "провести" (sci.), "выполнить" (off.-bus.)
- "Функционировать" → "работать" (all registers)
- "Обладать" → a concrete verb based on meaning
- Straighten "не X, а Y" — the only technique that works in ALL registers. This is a pure AI pattern. "Важно не количество, а качество" → "Важно качество".
- Kill banal metaphors everywhere. "Как дом без фундамента" reveals AI in any register.
- Add specifics (numbers, names, dates, titles) — a universal technique. But do not invent: if exact data is unavailable, leave the general formulation.
Critical restrictions by register:
- Научный: NO parcellation, rhetorical questions, particles ("ну", "вот", "ведь", "-то"), idioms, personal "я". Passive voice and "является" are normal. Authorial "мы" is allowed.
- Официально-деловой: NO conversational elements of any kind, personal voice, rhetorical questions, parcellation. Genre clichés ("в соответствии с", "на основании") MUST BE PRESERVED — these are not AI stamps.
- Разговорный/Публицистический: particles, idioms, personal voice allowed — but dosed (≤1 particle per paragraph, ≤1 idiom per 500–1000 characters).
Step 3 — Final verification
After editing, run through the checklist. The full list with register-dependent criteria and anti-patterns is in references/checklist.md. Load this file before verification.
Universal criteria (all registers):
- The text contains none of the markers from the lexical cliché catalog
- Template contrasts "не X, а Y" have been removed
- Banal metaphors have been removed or replaced
- The meaning of the source text is preserved
- No fabricated facts — all added specifics are real
- English calques are eliminated
- AI bureaucratese removed: "является" as a copula (except Научный and Официально-деловой), "представляет собой" (except Научный and Официально-деловой)
Mandatory overcorrection test: read the text and ask — "Could this text have been written by a human in this genre?" If a scientific article sounds like a TikTok post — the register is violated, roll back conversational edits. If a scientific or business text contains "ну", "ведь", "-то", idioms, or an inserted "я" — remove immediately.
Key principles
- Do not harm meaning. Style is secondary to content.
- Register is law. Do NOT lower the register. A scientific text must not become conversational.
- Do not overdo it. Conversational elements work on contrast: ≤1 particle per paragraph in publicist text, 0 in scientific and business.
- Verify facts. Better to leave a general formulation than to invent a number.
- Imitate the author, not the "average speaker." Lively Russian in a scientific article is not the same as in a Telegram chat.
- Organic unevenness > mechanical evenness. Human text is characterized by uneven depth and stylistic variation, not a collection of "human-like" wordlets.
- If in doubt — don't change. Conservative editing is better than aggressive transformation.
Anti-patterns (what not to do)
Full catalog is in references/checklist.md.
In brief:
- ❌ Lowering the register (Научный → Разговорный).
- ❌ Inserting "я" into formal text.
- ❌ Spamming conversational words in any register.
- ❌ Adding idioms to scientific and business text.
- ❌ Changing "является" to "это" in scientific and business style (it's normal there).
- ❌ Removing genre-specific bureaucratese in Официально-деловой text.
- ❌ Creating "mechanical emotions" — inserting "я считаю" without semantic weight.
- ❌ Turning every paragraph into a rhetorical question.
- ❌ Replacing one uniformity with another (all sentences became short or all paragraphs start with a particle).
Files (skills)
-
references
-
ai-markers.md 10.6 KB
Loading… -
checklist.md 5.6 KB
Loading… -
transformations.md 16.4 KB
Loading…
-
-
README.md 4.4 KB
# Humanize RU A standalone [Agent Skill](https://agentskills.io/specification) for text transformation and post-editing of AI-generated content. It removes signs of machine generation from Russian-language AI text while preserving its original register and communicative purpose — scientific text stays scientific, business text stays business, conversational stays conversational. Only the "AI smell" is removed. ## When to use | Situation | Use this skill | | ----------------------------------------------------------------------------------------------- | -------------- | | Text from an LLM (ChatGPT, Claude, Gemini, …) needs to be published under human authorship | ✅ | | A client or instructor checks the text with an AI detector | ✅ | | Machine output sounds dry, formulaic, or unnatural for its genre | ✅ | | AI output needs to be adapted to a specific tone of voice in Russian | ✅ | ## How it works The skill follows a four-step procedure built around one core principle: **not "make the text conversational", but "remove AI markers, keeping the register appropriate for the context."** ``` Step 0 Step 1 Step 2 Step 3 Determine ──► Diagnose ──► Transform ──► Verify the register AI markers (respecting (overcorrection the register) test) ``` ### Step 0 — Determine the register The appropriate transformations depend entirely on the functional style: | Register | Genre examples | | ------------------------- | -------------------------------------------------------- | | **Разговорный** | Telegram post, personal letter, comment | | **Публицистический** | Blog, media article, column, longread | | **Научный** | Scientific article, dissertation, textbook, report | | **Официально-деловой** | Order, contract, business letter, instruction | | **Художественный** | Story, novel, essay, literary sketch | **Register is law.** Never lower the register without an explicit reason — a scientific text "humanized" to conversational level is a ruined scientific text. ### Steps 1–3 — Diagnose, transform, verify 1. **Diagnose** — explicitly list every AI marker found. Transformation does not begin until the markers are enumerated. 2. **Transform** — apply techniques *selectively*, based on which markers were found and which register was determined. Transformations must not change the register. 3. **Verify** — apply the **mandatory overcorrection test**: "Could this text have been written by a human in this genre?" If a scientific article sounds like a TikTok post, roll back. ## Key principles - **Do not harm meaning.** Style is secondary to content. - **Register is law.** Do not lower the register. - **Do not overdo it.** Conversational elements work on contrast: ≤1 particle per paragraph in publicist text, 0 in scientific and business. - **Verify facts.** Better to leave a general formulation than invent a number. - **Organic unevenness > mechanical evenness.** ## Bundled resources | File | Purpose | | ------------------------------- | -------------------------------------------------------------------- | | `SKILL.md` | Skill definition, the four-step procedure, and universal rules | | `references/ai-markers.md` | Full catalog of AI markers with examples and explanations | | `references/transformations.md` | Transformation techniques with dosage tables per register | | `references/checklist.md` | Verification checklist with register-dependent criteria | ## Compatibility The skill is agent-agnostic. It describes **what** to do, not which specific tools to call, so it works with any AI agent that can read and write text. No external services are required. The skill follows the [Agent Skills specification v1](https://agentskills.io/specification). ## License MIT -
SKILL.md 11.1 KB
--- name: humanize-ru description: > Removes signs of machine generation from Russian-language AI text, preserving its original register and communicative purpose. Does not turn everything into conversational style — scientific text remains scientific, business text remains business, and conversational remains conversational. Use when you need to remove AI markers from text for publication, bypass AI detectors, or adapt machine output to natural Russian speech in the required register. --- # Humanize RU — removing AI markers from Russian text Remove signs of machine generation from the text, preserving its original register and communicative purpose. **Core principle:** not "make the text conversational", but "remove AI markers, keeping the register appropriate for the context." ## When to use - Text from ChatGPT, Claude, Gemini, or another LLM needs to be published under human authorship. - A client or instructor checks the text with an AI detector. - Machine text sounds dry, formulaic, unnatural for its genre. - AI output needs to be adapted to a specific tone of voice. ## Procedure ### Step 0 — Determine the text register FIRST, determine the functional style (register) of the source text. The appropriate transformations — and the ones that would ruin the text — depend on this. **Five registers and their key features:** | Register | Lexical features | Syntax | "I" allowed? | Genre examples | | ---------------------- | ------------------------------------------------------------------------- | ------------------------------------------------------------------- | ------------------------------------- | --------------------------------------------------------- | | **Разговорный** | Conversational vocabulary, particles, contractions, ellipsis | Short sentences, parcellation, inversions | Yes, freely | Пост в Telegram, личное письмо, комментарий | | **Публицистический** | Common vocabulary + expressive elements, rhetorical devices | Varied: alternating short and long, questions, exclamations | Yes, author's position | Блог, статья в СМИ, колонка, лонгрид | | **Научный** | Terms, abstract vocabulary, verbal nouns | Complex syntax, passive voice, clear logical structure | Only "мы" (author's) in methodology | Научная статья, диссертация, учебник, аналитический отчёт | | **Официально-деловой** | Bureaucratic formulas, business clichés, standardized phrasing | Cumbersome, clipped, unemotional, strict word order | No | Приказ, договор, деловое письмо, инструкция, отчёт | | **Художественный** | Figurative vocabulary, metaphors, individual-authorial words | Free, subordinated to artistic purpose | Depends on narrator | Рассказ, роман, эссе, художественный очерк | **How to determine the register** (check in order, first match is the answer): 1. **Formal marker?** Приказ, договор, инструкция, официальное письмо → Официально-деловой. 2. **Scientific vocabulary?** Terms, citations, methodology, abstract concepts as subjects → Научный. 3. **Literary devices?** Plot, characters, explicit imagery, metaphors as foundation → Художественный. 4. **Personal voice + expression?** "Я", rhetorical questions, value judgments, calls to action — without rigid scientific structure → Публицистический. 5. **Short, informal, with particles?** "Ну", "вот", "короче", ellipsis, emoji → Разговорный. **If unsure — default to Публицистический.** This is the neutral register, from which you can shift in any direction at the user's request. **NEVER lower the register without an explicit reason.** A scientific text "humanized" down to a conversational level is a ruined scientific text. ### Step 1 — Diagnosis: find AI markers Find and explicitly list all signs of machine origin in the text. The full marker catalog with examples and explanations is in [references/ai-markers.md](references/ai-markers.md). Load this file before starting diagnosis. Brief overview of marker categories: - **Lexical clichés** — formulaic phrases: "в современном мире", "следует отметить", "играет важную роль", "представляет собой" and others. - **Structural patterns** — uniform paragraph openings, "rule of three", identical sentence length, template contrasts "не X, а Y". - **Syntax** — monotonous word order, genitive chaining, overuse of "который", excessive passive voice. - **Style** — facelessness, emotional neutrality, encyclopedic style, absence of conversational elements, inappropriate metaphors. - **English calques** — "является" instead of zero copula, "делать" as a universal verb, mandatory subject, "иметь" instead of specific verbs. - **Content signs** — lack of specifics, logical loops, absence of context, hallucinations. **CRITICAL: do not proceed to transformation without listing the found markers explicitly.** Output the result in the following format: ``` ### Diagnostic Result **Lexical clichés:** [list of found markers or "none found"] **Structural patterns:** [list: "rule of three", "не X, а Y", uniform paragraph openings, ...] **Syntax:** [list: monotonous sentence length, genitive chaining, ...] **Style:** [list: flat tone, explanatory intonation, absence of personality, ...] **English calques:** [list or "none found"] **Content signs:** [list: lack of specifics, logical loops, ...] **Total markers:** N (3+ = text almost certainly machine-generated) ``` ### Step 2 — Transformation: removing AI markers while respecting the register Apply techniques SELECTIVELY: which markers were found + which register was determined. **Key rule:** transformation must NOT change the register. Scientific text remains scientific, conversational remains conversational. Remove only the unnaturalness of AI generation, not the genre traits of the register. The full catalog of techniques with dosage tables by register is in [references/transformations.md](references/transformations.md). Load this file before transformation. Below are the critically important rules that must not be violated: **Universal rules (all registers):** - "Осуществить" → "сделать" (conv.), "провести" (sci.), "выполнить" (off.-bus.) - "Функционировать" → "работать" (all registers) - "Обладать" → a concrete verb based on meaning - **Straighten "не X, а Y"** — the only technique that works in ALL registers. This is a pure AI pattern. "Важно не количество, а качество" → "Важно качество". - **Kill banal metaphors** everywhere. "Как дом без фундамента" reveals AI in any register. - **Add specifics** (numbers, names, dates, titles) — a universal technique. But do not invent: if exact data is unavailable, leave the general formulation. **Critical restrictions by register:** - **Научный:** NO parcellation, rhetorical questions, particles ("ну", "вот", "ведь", "-то"), idioms, personal "я". Passive voice and "является" are normal. Authorial "мы" is allowed. - **Официально-деловой:** NO conversational elements of any kind, personal voice, rhetorical questions, parcellation. Genre clichés ("в соответствии с", "на основании") MUST BE PRESERVED — these are not AI stamps. - **Разговорный/Публицистический:** particles, idioms, personal voice allowed — but dosed (≤1 particle per paragraph, ≤1 idiom per 500–1000 characters). ### Step 3 — Final verification After editing, run through the checklist. The full list with register-dependent criteria and anti-patterns is in [references/checklist.md](references/checklist.md). Load this file before verification. **Universal criteria (all registers):** - [ ] The text contains none of the markers from the lexical cliché catalog - [ ] Template contrasts "не X, а Y" have been removed - [ ] Banal metaphors have been removed or replaced - [ ] The meaning of the source text is preserved - [ ] No fabricated facts — all added specifics are real - [ ] English calques are eliminated - [ ] AI bureaucratese removed: "является" as a copula (except Научный and Официально-деловой), "представляет собой" (except Научный and Официально-деловой) **Mandatory overcorrection test:** read the text and ask — "Could this text have been written by a human in this genre?" If a scientific article sounds like a TikTok post — the register is violated, roll back conversational edits. If a scientific or business text contains "ну", "ведь", "-то", idioms, or an inserted "я" — remove immediately. ## Key principles 1. **Do not harm meaning.** Style is secondary to content. 2. **Register is law.** Do NOT lower the register. A scientific text must not become conversational. 3. **Do not overdo it.** Conversational elements work on contrast: ≤1 particle per paragraph in publicist text, 0 in scientific and business. 4. **Verify facts.** Better to leave a general formulation than to invent a number. 5. **Imitate the author, not the "average speaker."** Lively Russian in a scientific article is not the same as in a Telegram chat. 6. **Organic unevenness > mechanical evenness.** Human text is characterized by uneven depth and stylistic variation, not a collection of "human-like" wordlets. 7. **If in doubt — don't change.** Conservative editing is better than aggressive transformation. ## Anti-patterns (what not to do) Full catalog is in [references/checklist.md](references/checklist.md). **In brief:** - ❌ Lowering the register (Научный → Разговорный). - ❌ Inserting "я" into formal text. - ❌ Spamming conversational words in any register. - ❌ Adding idioms to scientific and business text. - ❌ Changing "является" to "это" in scientific and business style (it's normal there). - ❌ Removing genre-specific bureaucratese in Официально-деловой text. - ❌ Creating "mechanical emotions" — inserting "я считаю" without semantic weight. - ❌ Turning every paragraph into a rhetorical question. - ❌ Replacing one uniformity with another (all sentences became short or all paragraphs start with a particle).
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.