Review the text provided in $ARGUMENTS (or the most recent Russian text output if no arguments) using the ru-text skill.

This skill reads the corpus that ships with the ru-text skill, which is installed alongside it. Locate that folder once, then read the named files from it:

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:

Then check the text: typography mechanically (straight quotes, a hyphen doing a dash's work, for an ellipsis, 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.

Read-only — do NOT modify the source files. 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, removes the listed tools from the pool while the command is active — verified on 2.1.220 by attempting a write through , through redirection, through , and through a delegated subagent: every path was denied and no file was created. alone would not achieve this; that field pre-approves tools, it does not restrict them.

The list removes the direct file-writers — , , — and the command-executing tools we identified and tested on 2.1.220: , and . is easy to miss: it runs commands too, and it follows permission rules, so a broad allow-rule silently pre-approves it while a 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 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 , 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: