cleanup-ai-noise

name	cleanup-ai-noise
description	Use at the end of any article-writing skill (habr-article, vcru-article, dzen-article, pikabu-article, telegraph-article, tenchat-article, telegram-article, telegram-announcement) before publication. Saves the draft to .tmp/current/articles/, invokes the ai-text-checker agent with platform context, presents the cleanup report and Needs-human-decision items to the author. Detective companion to the preventive skill `living-text-style` (invoked at ФАЗА 0 before drafting).
allowed-tools	Read, Write, Bash, Task

Cleanup AI Noise

Trigger skill that wires an article-writing flow into the ai-text-checker agent. The agent owns the rules; this skill owns the plumbing.

When to Use

Invoke from ФАЗА 3 of any article-writing skill, BEFORE the manual «Не робот» checklist. The article-writing skill produces the draft text in conversation; this skill persists it, runs the agent, and presents the result.

Do NOT invoke for non-article content.

Inputs

• platform (required): one of habr | vc | dzen | pikabu | telegraph | tenchat | telegram | telegram-announcement
• The final draft text from ФАЗА 2 of the calling article-skill (it lives in conversation context)

Steps

1. Generate filename and path

filename = draft-{platform}-{YYYYMMDD-HHmm}.md
path     = .tmp/current/articles/{filename}

Use date +%Y%m%d-%H%M via Bash for the timestamp (Moscow time is fine — the file is throwaway).

2. Ensure target directory exists

mkdir -p .tmp/current/articles

3. Save the draft

Write the full final article text from ФАЗА 2 to path using the Write tool. Do not modify the text — pass it through verbatim.

4. Invoke the agent

Call the Task tool with:

subagent_type: ai-text-checker
description:   AI-noise cleanup for {platform} draft
prompt:        |
  file_path: .tmp/current/articles/{filename}
  platform: {platform}

  Run the full ai-text-checker workflow on this draft.

Wait for the agent to return.

5. Read the cleaned file

After the agent returns, Read the file at path. The file is now the post-cleanup version (the agent edited in-place).

6. Present results to the author

Output to the user a Russian-language summary. The agent prints the actual backup path in its report — pass it through verbatim.

Чистка AI-шума завершена.

Файл (после чистки): {file_path the agent worked on}
Backup (исходник):   {BACKUP_PATH from the agent's report — under .tmp/article-backups/{platform}/}

{полный текст отчёта от агента — таблицы статистики, ключевые правки, спорные места}

Дальше — ручной чек-лист «Не робот» из ФАЗА 3 текущего article-скилла.
Если в блоке «Требует решения автора» есть пункты — реши их вручную перед публикацией.

Output

The author sees:

• Path to cleaned text + backup (centralized in .tmp/article-backups/{platform}/)
• Agent's structured report (metrics, layer statistics, top edits, items needing human decision)
• Reminder to run the manual «Не робот» checklist

Backup convention (IMPORTANT)

Backups are NEVER stored next to the source article. The ai-text-checker agent writes them to:

.tmp/article-backups/{platform}/{stem}-{YYYYMMDD-HHmm}.bak.md

Example: .tmp/article-backups/habr/codex-128-goal-experiments-20260504-1947.bak.md

This keeps articles/{platform}/ directories clean (only finished pieces, no .bak clutter). Since .tmp/ is git-ignored, backups don't pollute commits either.

Notes

• The ai-text-checker agent is the Single Source of Truth for the AI-marker catalog. This skill must NOT duplicate any rules from the agent body.
• If the calling article-skill already has a different draft persistence convention (e.g., a final articles/{platform}/{slug}.md save), use that path INSTEAD of .tmp/current/articles/ — the agent works on whatever path is passed in. Backups still go to the centralized location regardless.
• Per-skill discovery: /cleanup-ai-noise triggers this skill in any session.