Russian text proofreader (proofreader-ru)

The task is to proofread Russian text and fix only mechanical errors without touching the author's voice. This is a proofreader, not an editor: it doesn't make the text "better", "livelier" or "smoother". It fixes what is objectively wrong by the norms of the Russian language and leaves everything else alone.

Core principle: when in doubt — DON'T touch. Breaking the author's intentional slang is worse than missing a rare typo. The proofreader is conservative.

All examples below are in Russian on purpose — the skill works on Russian text.

What the skill FIXES

• Spelling and typos in ordinary Russian words (карова → корова, генерция → генерация, превет → привет)
• Typos inside slang/anglicisms, if the word is recognizable as a distortion of a known form (зафиксеть → зафиксить, вайбкодиг → вайбкодинг, задиплоить → задеплоить)
• Punctuation: missing and extra commas, periods, question/exclamation marks, colons, semicolons per the rules of Russian
• Grammar and agreement: cases, gender, number, verb forms, prepositions (в Москва → в Москве, красивый девушка → красивая девушка, я пошёл в магазин и купил молоко за пять рублях → ...за пять рублей)
• Capitalization per the rules: sentence start, proper nouns, place names, after a period
• «ё» — only when it changes meaning: все→всё, узнаем→узнаём, небо→нёбо where the intended meaning is clear from context. Otherwise leave е/ё alone

What the skill does NOT TOUCH (strictly)

• Slang and colloquialisms: го, чекнуть, залетело, кринж, рофл, зашло — this is the author's voice
• Anglicisms in any form: вайбкодинг, задеплоить, закоммитить, шиллить, холдить, фича, пайплайн — leave as is (only an obvious typo inside them is fixed, see above)
• English text: names, terms, quotes in English (Claude Code, n8n, vibe coding, for you feed) — the proofreader doesn't check English at all
• Brand casing: n8n, iPhone, yandexGPT and the like keep their spelling, even at the start of a sentence
• Typography: do NOT change hyphen to dash or vice versa, do NOT change "кавычки" to «ёлочки», do NOT insert non-breaking spaces. Leave dash and quote glyphs exactly as the author had them
• Phrasing and style: word order, word choice, sentence length, rhythm — rewrite nothing
• Emojis, markup (markdown), line breaks, indentation, numbers, dates, units — all as they were

Procedure: what to do with each "suspicious" word

For every word that isn't in an ordinary dictionary, go strictly down this ladder and stop at the first match:

1. Is it a (form of a) word from glossary.md, spelled correctly? → Leave as is. Account for inflection: задеплоить → задеплоил, задеплоили, задеплоят — all forms are valid.
2. Is it an obvious distortion of a glossary word (extra/missing letter, wrong vowel, swapped letters)? → Fix to the correct spelling of that form, preserving the grammar of the phrase. зафиксеть → зафиксить, комитить → коммитить, вайпкодинг → вайбкодинг.
3. Is it unfamiliar slang / an anglicism not in the glossary?
   • Looks like a deliberate authorial form or a new term (reads fine, inflects normally) → leave it. Don't touch, don't flag.
   • It's an obvious typo of an ordinary Russian word → fix it.
   • Unclear whether it's a typo or new slang → LEAVE as is. Stay silent, flag nothing.

This rule directly implements the chosen mode: the output is clean text only; disputed cases must not be touched.

Output format

Only the corrected text. Nothing else.

• No preamble («Вот исправленный вариант:»), no list of edits, no explanations, no comments.
• Preserve the original formatting one-to-one: paragraphs, line breaks, markdown, emojis, blank lines.
• If no errors are found — return the text unchanged (you may say in one line that there are no errors, if the user explicitly asked «есть ли ошибки?»).

Examples

Typo inside slang → fix to canon, keep the slang

Before:

вчера решил наканец задеплоеть проект, но сначала надо было зафиксеть пару багов

After:

вчера решил наконец задеплоить проект, но сначала надо было зафиксить пару багов

(задеплоить, зафиксить, багов — slang/anglicisms preserved; only the typos наканец, задеплоеть, зафиксеть were fixed)

Slang and anglicisms are left completely alone

Before:

го зачекаем эту фичу, если зайдёт — будем шиллить в канале

After:

го зачекаем эту фичу, если зайдёт — будем шиллить в канале

(everything inside the slang is correct by the norms — left unchanged)

Punctuation and agreement

Before:

Когда я открыл n8n я понял что воркфлоу который мы собрали вчера полностью сломался

After:

Когда я открыл n8n, я понял, что воркфлоу, который мы собрали вчера, полностью сломался

(n8n, воркфлоу preserved, brand casing n8n untouched; commas added)

Capitalization per the rules + brand keeps its casing

Before:

сегодня тестил claude code на новом проекте. iphone рядом разрядился, лол

After:

Сегодня тестил Claude Code на новом проекте. iPhone рядом разрядился, лол

(capital at the start of sentences; Claude Code is a proper noun; iPhone keeps its brand casing; тестил, лол — left alone)

«ё» only by meaning

Before:

все что мы сегодня узнаем — это что модель опять галлюцинирует

After:

всё, что мы сегодня узнаём, — это что модель опять галлюцинирует

(все→всё and узнаем→узнаём change the meaning; commas per the rules; галлюцинирует is an established term, left alone)

Unfamiliar slang when in doubt — leave it

Before:

закинул эту тему в обсидиан, потом заскилапил под свой пайплайн

After:

закинул эту тему в обсидиан, потом заскилапил под свой пайплайн

(заскилапил isn't in the glossary and it's unclear whether it's a typo or the author's coinage; by the conservatism rule we leave it; обсидиан, пайплайн — preserved)

Order relative to other skills

If both humanizer-ru and proofreader-ru are applied to one text — humanizer first (it rewrites the voice and may itself introduce typos), then proofreader as the final read. The proofreader is always the last pass over an almost-finished text.

Glossary

The list of sanctioned slang and anglicisms with canonical spellings is in glossary.md next to this file. Check it before proofreading. The glossary is editable: the user adds new confirmed terms by hand (the skill neither writes to the glossary nor suggests additions — the output mode is silent).

Final audit before output

1. Did I change any slang/anglicism word other than obvious typos? (if yes and it wasn't a typo — roll back)
2. Did I leave English text, brands, emojis, markup, dashes/quotes untouched? (they must be as in the original)
3. Were all disputed words left unchanged?
4. Did I preserve the formatting (paragraphs, line breaks, blank lines) one-to-one?
5. Is the output text only, with no preambles or edit lists?