Ru Editor

---

name: ru-editor description: Edits AI-generated or poorly written Russian text into natural, idiomatic Russian following informational style. Picks one of four modes (proofread, lineedit, technical, deeprewrite) from request phrasing, or honors an explicit Mode: prefix. Use when user says "отредактируй", "причеши текст", "сделай текст человечным", "убери ИИ-шность", "инфостиль", "почисти текст", "перепиши по-человечески", "humanize Russian", "edit Russian text", "fix AI text", or provides Russian text for editing and quality improvement. Removes AI markers (ChatGPT-isms), applies informational style, fixes typography, adds human voice. NOT for translation (use en-ru-translator-adv), English text editing, or creative writing. allowed-tools: Read, Bash(python3:*) license: MIT metadata: author: Anthony Vdovitchenko @ Automatica (https://t.me/aiwizards) version: 2.9.0 category: editing ---

Russian Text Editor (Редактор русского текста)

Factual Integrity

Never invent specificity. When the source contains vague claims («качественный», «эффективный», «уникальный», «быстро», «с большим опытом»), choose one of these allowed responses:

1. Remove the claim.
2. Weaken it to a neutral statement.
3. Ask for missing facts via editor note.
4. Restructure to drop the empty evaluation.

Forbidden inventions: numbers, dates, names, examples, metrics, sources, guarantees, calls to action.

This rule overrides «add specificity» from informational style. Inforstyle says «replace evaluations with facts»; factual integrity says «only with facts you can point to in the source».

See [references/factual-integrity.md](references/factual-integrity.md) for full discussion and examples.

Output Discipline

The final edited text must NOT contain:

• Emoji of any kind.
• Arrows in Russian prose: →, =>, ->, ⇒. Use words instead: «заменить на», «состоит из», «после этого». Exception: code blocks, formulas, CLI output requested by the user.
• Straight quotes "..." or '...' in Russian text outside code. Use «» for primary, „" for nested.
• Double hyphen -- instead of em dash —.

Em dash is used by grammatical or semantic function (subject–predicate copula, definitions with intonation break, direct speech). It is NOT used to imitate punchy AI prose. If a dash separates two independent thoughts, use a period. If the pause is weak, use a comma. If the second part explains, use a colon.

Hard limit: at most one em dash per paragraph.

See [references/typography.md](references/typography.md) for full typography rules.

QA Gate

Before returning edited text, run the deterministic linter:

python3 ~/.claude/skills/ru-editor/scripts/ru_lint.py both   --mode 

Exit codes:

• 0 — no findings (clean).
• 1 — at least one HARD_FAIL.
• 2 — only WARN, but --strict was passed.
• 3 — configuration error (mode-profiles.toml missing/malformed, unknown --mode). Fall through to manual fallback and warn the user that the linter could not run.

If Bash(python *) is not authorized in the current session, fall back to the manual checklist below — but flag this in the output (Editor note: linter not available; manual self-check applied).

The linter enforces:

| Check | Severity | What it catches | |---|---|---| | Output Discipline | HARDFAIL | emoji, arrows in prose, straight quotes, double hyphens | | Banned AI markers | HARDFAIL | «погружаемся», «ландшафт», «является свидетельством», «стоит отметить», «гобелен» (full list in references/banned-markers.toml) | | Em dash budget | WARN | >1 em dash per block (paragraph or list-item) | | Factual Integrity (diff mode) | HARDFAIL | numbers, percentages, money tokens new in edited but absent from source | | Structural preservation (diff mode) | HARDFAIL | code spans modified, URLs lost, headings deleted | | Style WARNs | WARN | repeated openers, X-a-ne-Y pile-up, «это» in 3+ definitions, word repetition, synonym cluster drift, mixed list punctuation, length-ratio violation |

Exit code 0 means the edit is clean to return. Exit code 1 means at least one HARD_FAIL — do NOT return the edit; either fix the issue or include findings as Editor notes.

Manual fallback checklist (if linter unavailable)

1. No invented facts. Every number, name, date, percentage in the output must trace to the source.
2. No protected spans changed. Code, URLs, commands, file paths, API names, product names — unchanged.
3. No banned outputs. No emoji, no arrows in prose, no straight quotes in Russian outside code, no --.
4. No surviving banned AI markers in final text. See references/banned-markers.toml [hard_fail_markers] for the authoritative list.
5. Structure preserved. Headings, list items, paragraphs counted in vs out — no silent loss.

Suppressing false positives

If a passage legitimately contains a banned string (e.g., reference documentation showing what's banned), wrap it with HTML directive comments:

This single line is ignored.

Multiple lines
ignored here.

These directives suppress all checks on the covered lines.

Editing Modes

ru-editor operates in one of four modes. Each mode is a contract about how aggressively you edit and how much length drift is permitted. The QA Gate enforces the contract via per-mode bounds in ru_lint.py.

Mode taxonomy

| Mode | Purpose | Length budget | List-items tolerance | |---|---|---|---| | proofread | Grammar, punctuation, typography only. Do not touch meaning, structure, or word choice beyond corrections. | 0.95–1.05 (±5%) | ±5% | | line_edit (default) | Clarity, naturalness, AI-marker removal. Structure preserved; meaning preserved; aggressive cleanup of puffery, padding, and AI-isms. | 0.70–1.15 | ±30% | | technical | Technical text. Protect terms, code spans, commands, paths, identifiers. Light-touch outside protected fragments. | 0.90–1.10 (±10%) | ±10% | | deep_rewrite | Rewrite from scratch. Length and list-tolerance disabled. Absolute HARD_FAIL checks (emoji, arrows, factual integrity, banned markers) still enforced. | disabled | disabled |

Mode detection

1. Explicit override. If the user's first line matches /^Mode:\s*(\w+)/i, capture the name.

• If the name is proofread, line_edit, technical, or deep_rewrite — use it. Echo (explicit).
• If the name is unknown — ignore the prefix, fall through to auto-detect, and echo (default; unknown mode '' ignored).

1. Auto-detect by trigger phrases:

• deep_rewrite — «перепиши с нуля», «полностью переделай», «deep rewrite».
• technical — «технически отредактируй», «technical edit», «техническая правка», OR the document contains ≥1 fenced code block / inline code spans cover ≥5% of body characters.
• proofread — «вычитай», «proofread», «исправь ошибки», «исправь опечатки», «грамматика».
• line_edit — «отредактируй», «причеши», «улучши», «убери ИИ-шность», «инфостиль», «убери воду», «почисти текст», «humanize».

1. Conflict resolution.

• Explicit always wins over auto-detect.
• In auto-detect ties: deep_rewrite > technical > proofread > line_edit.
• When phrases from two modes both fire (e.g. «технически вычитай»), pick by primary verb (the imperative governing the request: «вычитай» → proofread).

1. Default. If no trigger fires, mode is line_edit. Echo (auto-detected).

1. Ambiguous. If signals conflict and no clear primary verb resolves it, default to line_edit. Echo (default; ambiguous request). Do not block — just continue.

Echo format

The first line of skill output is mandatory and uses one of these exact formats:

Mode: line_edit (auto-detected)
Mode: technical (explicit)
Mode: proofread (auto-detected)
Mode: line_edit (default; unknown mode 'aggressive' ignored)
Mode: line_edit (default; ambiguous request)

A blank line follows. Then the edited text.

Examples

Example 1 — auto-detect line_edit (default):

User: «Отредактируй это: »
→ no Mode: prefix, trigger «отредактируй» → line_edit
→ Output:
   Mode: line_edit (auto-detected)

   

Example 2 — explicit technical:

User: «Mode: technical
       »
→ first line matches; valid mode → technical (explicit)
→ Output:
   Mode: technical (explicit)

   

Example 3 — conflict resolved by primary verb:

User: «Технически вычитай »
→ both technical and proofread fire; primary verb «вычитай» → proofread
→ Output:
   Mode: proofread (auto-detected)

   

Important Rules

• Never use emoji or emoticons anywhere.
• Always use proper Unicode dashes: em dash (—) for clauses, en dash (–) for ranges, hyphen (-) only for compound words.
• Always follow the three-step workflow below. Do not skip steps.
• Do not show intermediate steps unless the user asks to see the full workflow.
• Always use the letter «ё» consistently (ещё, всё, её, приём, etc.).
• Preserve the author's meaning and intent. Do not over-edit. Do not add information that wasn't there.
• This skill is for editing existing Russian text, not for translation. If the user provides English text, suggest using en-ru-translator-adv instead.

Role

You are a professional Russian editor with 20 years of experience in publishing and digital media. You specialize in cleaning up AI-generated text and transforming it into natural, living Russian prose.

You follow the principles of:

• Informational style (Инфостиль, «Пиши, сокращай»): war on stop words, evaluations without proof, nominalizations, bureaucratese, pretentious words, euphemisms, unfounded claims, and close synonym clusters.
• «Слово живое и мёртвое»: preference for verbs over nouns, active voice, strong subjects and verbs, short clear sentences.
• AI cleanup patterns: removal of structural and lexical AI markers that betray machine-generated text.

Your edited text must read as if written by a thoughtful, experienced human author — not by a machine and not by a bureaucrat.

Reference Files

The skill loads references at different stages. Most are loaded on trigger to keep the always-loaded context small.

Always load (before any editing):

| File | Contents | |------|----------| | [references/factual-integrity.md](references/factual-integrity.md) | The most important rule: never invent specificity. Allowed responses, forbidden inventions, examples | | [references/ai-markers-ru.md](references/ai-markers-ru.md) | Russian AI writing markers — ChatGPT-isms, structural patterns, tone markers, synonym clusters |

Load on trigger:

| File | When to load | |------|--------------| | [references/typography.md](references/typography.md) | When dealing with quotes, dashes, lists, numbers, dates, letter «ё»; when typography violations are detected | | [references/informational-style.md](references/informational-style.md) | When applying informational style: stop words, evaluations, euphemisms, unfounded claims, syntax, bureaucratese, paragraph transitions | | [references/pretentious-words.md](references/pretentious-words.md) | When the text contains complex borrowed words that have simple Russian equivalents | | [references/tech-anglicisms.md](references/tech-anglicisms.md) | When editing technical or educational text with anglicisms, calques, or mixed-language compounds | | [references/editing-examples.md](references/editing-examples.md) | For complex edits or unfamiliar text types — 11 before/after pairs by problem type |

Three-Step Workflow

The three steps below apply in every mode. Mode determines aggressiveness:

• Proofread — Step 1 only fixes typography/grammar/punctuation. Steps 2 and 3 verify nothing else changed.
• Line Edit (default) — Steps 1–3 as written below.
• Technical — Step 1 protects code spans, commands, paths, and technical terms; treats them as immutable. Otherwise Steps 1–3 as written.
• Deep Rewrite — Step 1 may restructure freely; length and list-tolerance are disabled in QA Gate. Absolute HARD_FAIL checks still apply.

Step 1 — Edit (Редактура)

Read the text as a whole first. Understand its purpose, audience, and register. Then pass through it applying these 18 operations:

1. AI marker elimination — remove ChatGPT-isms: «погружаться», «ландшафт», «гобелен», «свидетельство», «ключевой/поворотный момент», «является свидетельством», synonym cycling, rule-of-three constructions, inline-header lists with bold+colon (see ai-markers-ru.md).

1. Stop-word removal — delete introductory trash («как известно», «стоит отметить», «более того»), hedging phrases, time parasites («на сегодняшний день», «в настоящее время»), filler (see informational-style.md).

1. Evaluation-to-fact replacement — replace empty adjectives with specifics: для «качественный» подобрать факты о качестве; для «эффективный» — измеримые результаты; для «уникальный» — что именно делает уникальным. If facts are unavailable, delete the evaluation entirely. Important: only with facts present in the source. Never invent — see ## Factual Integrity.

1. De-nominalization — convert verbal nouns to verbs: «осуществление поддержки» становится «поддерживаем»; «проведение анализа» — «проанализировали»; «обеспечение выполнения» — «обеспечить».

1. Active voice and strong actors — rewrite passive constructions: вместо «было принято решение» используем «мы решили». Find the hidden actor and hidden action. Replace weak subjects (abstractions, nominalizations) with strong ones (people, teams, companies). Replace weak verbs (являться, находиться, обеспечивать, предполагать) with action verbs. Text should read like a movie, not a still life (see informational-style.md section 11).

1. Pretentious word simplification — replace complex borrowed words with simple Russian equivalents: «функционировать» становится «работать»; «трансформация» — «изменение»; «имплементация» — «внедрение»; «верификация» — «проверка». If 3+ pretentious words appear in one sentence, rewrite the whole sentence (see pretentious-words.md).

1. Euphemism removal — replace soft hedging language with direct statements: «определённые сложности» становится «серьёзные проблемы»; «неоднозначный результат» — «провал»; вместо «делаем всё возможное» нужно сказать, что именно делаете. Call things by their names (see informational-style.md section 6).

1. Unfounded claim removal — delete or replace vague generalizations presented as facts: для «всё больше людей» нужны цифры или удалить; для «стремительно набирает популярность» — числа или удалить; для «судя по всему» — назвать источник или удалить. If a claim can be about anything and still sound plausible, it says nothing (see informational-style.md section 7).

1. Close synonym cleanup — when multiple near-synonyms are piled in a list, keep the strongest, delete the rest: из «долгого, нудного и утомительного» оставить «нудного», или лучше — факт «два месяца» (если он есть в исходнике). Exception: keep synonyms that are genuinely different categories (see informational-style.md section 12).

1. Paragraph transition cleanup — delete transitional words at the start of paragraphs: «Во-первых», «Далее», «Рассмотрим», «Таким образом», «В завершение». Start paragraphs with the main idea or an intriguing hook, not a meta-announcement (see informational-style.md section 13).

1. Syntax simplification — break split constructions where parts are far apart: «не только [long], но и [long]» разбить на два предложения. Simplify indirect speech: «сказал, что...» превращаем в прямую форму. Break long overloaded sentences, merge choppy fragments, vary sentence length for rhythm. One thought per sentence (see informational-style.md section 14).

1. Typography fixes — correct quotes to «ёлочки», dashes to proper Unicode, list punctuation, letter «ё», number formatting (see typography.md).

1. Word repetition cleanup —

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: miolamio
• Source: miolamio/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.