Look for a directory named references whose parent directory is named ru-text, and which contains info-style.md. Every file named below sits in that folder.
Search with file tools, never with a shell. Glob on a pattern like **/ru-text/references/info-style.md, or whatever file search the host offers, then Read. This command has no use for a command line: Bash is on its disallowed-tools list, and hosts that do not implement that field still refuse it — measured in Claude Cowork on 12.08.2026, where an opening ls returned «Permission to use Bash has been denied» in red before the search found the corpus anyway. The call bought nothing and cost the reader a scare.
In Claude Code the folder is ${CLAUDE_SKILL_DIR}/../ru-text/references/ — Read info-style.md there first and skip the search. If that path arrives unexpanded or the read fails, fall back to the search above.
Do not guess a path. If the folder cannot be found, say so and stop — a check run against remembered rules instead of the corpus is not this command, and reporting one as the other is the failure this whole product exists to prevent.

Two depths, one command

This check runs at one of two depths. Nobody chooses a depth; these rules do.

Full — the «Check order» below, whole. Run it whenever a person asked for this check («вычитай», «прогони ru-text», a gate that names ru-text, a golden-set run) or triage escalated. An explicit request is never answered with triage.

Triage — for self-initiated runs only: you produced or encountered Russian text and are checking it out of discipline, with no instruction naming ru-text. Load three things and nothing else:

The ru-text SKILL.md — its inline typography table and top stop-words. Skip if it is already in context, which on an always-on host it usually is.
The «Neuroslop index» section of addenda.md: from the ## Neuroslop index heading to the next ##. Not the rest of the file — the rest is ten times the size.
Section «B. Каталог стоп-слов» of info-style.md: from its ## B. heading to the next ##. Not the rest of the file.

Then check the text: typography mechanically (straight quotes, a hyphen doing a dash's work, ... for an ellipsis, a space between a number and %, an ordinary space after в, к, с, о, у, и, а, я — verify by codepoint, not by eye); catalog stop-words including inflected forms, judging every candidate line yourself («данные» the noun is not «данный» the stop-word); index tells by eye.

Triage may report only what a single line decides: typography and confirmed catalog hits. A neuroslop tell is never a triage finding. Every AD rule carries carve-outs that live only in the full file, and «не X, а Y» with a real antecedent is ordinary prose — flagging it from an index alone is the false positive this command would lose the most trust for. A tell seen in triage is an escalation trigger and nothing else.

Escalate to full when any one of these holds: a neuroslop candidate appeared · five findings are confirmed · the text is bound for a reader (a deliverable, a publication, a client). Escalating is silent — continue into the full procedure as though it had been asked for.

A triage report names itself: «Быстрая проверка: типографика и стоп-слова. Полная вычитка по корпусу не выполнялась.» Reporting triage as the full check is the failure this product exists to prevent. No search tool on this host → no triage: run the full check.

Check order

Typography — read typography.md, then apply:
- Quotes: «» primary, „“ nested
- Dashes: — (em) in text, – (en) in ranges, - (hyphen) in compounds only
- Spaces: NBSP after single-letter words, in digit groups, before units — but % is glued to the number in general and web text (R37: 100%, not 100 %; a GOST technical document may use 100<NBSP>%). Report it with the reason the rule gives: an ordinary space is a line break point and leaves % alone on the next line
- Ellipsis, abbreviations, special characters
Anti-patterns — read anti-patterns.md, then scan for:
- Bureaucratic language and nominalization
- Passive voice overuse
- Sentence bloat
- Tautology and pleonasm
Writing quality — read info-style.md, then apply:
- Stop-words and filler
- Specificity and facts
- Structure and clarity
Domain-specific — load if text type is identifiable:
- UI/interface text → ux-writing.md
- Email/business → business-writing.md
- Needs grammar review → editorial-punctuation.md + editorial-grammar.md
Experience-based / neuroslop — read addenda.md, then scan for the AI-generated-prose tells:
- Manufactured antithesis (AD-6) — «не X, а Y» / «не просто X, а Y» with no antecedent
- Preemptive virtue qualifier (AD-7) — «без воды», «чётко, по делу»
- Assistant-register meta-commentary (AD-8) — «Отличный вопрос!», «Надеюсь, это помогло»
- Hollow openers (AD-9) — «давайте разберёмся», «погрузимся», «важно понимать, что»
- Declared sincerity (AD-10) — «честный разбор», «давайте будем честны»: honesty predicated of the piece. The same reflex on a single statement — «скажу честно: дедлайн сорван» — is AD-7
- Mandatory tricolon (AD-11) — «инновационный, трансформирующий, прорывной»
- Hollowed mechanism (AD-12) — «зависит от различных факторов», «свои особенности»
- Phantom attribution (AD-13) — «исследования показывают», «эксперты отмечают»
- Chat transcript as the artifact (AD-14) — the document's skeleton is a dialogue
- Search-engine addressee (AD-15) — the query phrase repeated where a pronoun would serve
- Additive pseudo-pair (AD-16) — «не только X, но и Y» where Y adds nothing
- Comma welded to a dash (AD-17) — a comma and an em dash side by side inside one sentence, both demanded by the same construction: «Отчёт, собранный за ночь, — на столе». Search by codepoint: the gap between the marks is normally the NBSP R16/R44 require before a dash, so a search written with an ordinary space finds nothing in correctly typeset text. Direct speech («Сроки поедут», — предупредила Петрова) is two constructions meeting, and a comma closing homogeneous subordinate clauses before the main clause forms a single mark with the dash — different grounds, same outcome: never this rule
- Uppercase band (AD-18) — three or more consecutive uppercase Cyrillic words of two letters or more, unbroken by a lowercase word or by a line break. Not a neuroslop tell, and it is absent from the index above on purpose: a person shouts on a keyboard with no italic key, a model does not. It belongs to this full pass and to nothing faster. One or two uppercase words are deliberate emphasis and are never flagged; abbreviations, machine text, status cells and headings do not count toward the run
The list above is a prompt for the eye, not the rule set. Two of these — AD-14 and AD-15 — are charged to the document, so ask them of the piece as a whole and not of any one sentence. Every rule has carve-outs that decide as many cases as the triggers do; they are in addenda.md, and a finding raised without checking them is the false positive this command costs the most trust for.

Output format

Read-only — do NOT modify the source files. ru-check is a check: it reports issues and returns the corrected text for the user to apply. It must not write to, edit, or overwrite the analysed file(s).

How this is enforced depends on the host, and it is worth being exact about it.

In Claude Code, disallowed-tools removes the listed tools from the pool while the command is active — verified on 2.1.220 by attempting a write through Write, through Bash redirection, through Monitor, and through a delegated subagent: every path was denied and no file was created. allowed-tools alone would not achieve this; that field pre-approves tools, it does not restrict them.

The list removes the direct file-writers — Write, Edit, NotebookEdit — and the command-executing tools we identified and tested on 2.1.220: Bash, PowerShell and Monitor. Monitor is easy to miss: it runs commands too, and it follows Bash permission rules, so a broad Bash allow-rule silently pre-approves it while a disallowed-tools: Bash entry does not remove it. The list names the paths we closed, not every path that exists.

Four limits, stated plainly. A connected MCP server can expose its own write tools, which a per-tool denial list cannot know about. A named subagent whose own definition grants it tools: Write is governed by that definition, not by a parent command's denial list. A later Claude Code release can add a tool this file does not name, and a hand-written list does not update itself. And on hosts that do not implement disallowed-tools, everything above is an instruction the agent is asked to follow, not a guarantee the platform enforces.

Where the mechanism ends, the contract does not: the rule is that this command returns text and never writes, on every host, whatever the tool roster happens to contain.

Returning the corrected text in the output rather than writing it keeps the command deterministic and free of side effects. Silently inserting NBSP into a source file is doubly harmful: a target that strips NBSP on import (e.g. Notion) shows the reader no change, and the same insertion breaks later exact-string tooling (grep/replace) on that file.

Return:

Corrected text
List of changes grouped by category (typography / style / grammar / domain)
Severity per change: critical / high / medium / low

ru-check

Russian Text Quality Check

Where the reference files live