En

English — Official English version of llm-text-hygiene.

LLM-Text-Hygiene — Remove AI residue from finished texts

Overview & Purpose

AI-assisted texts accumulate residue that remains invisible in draft form and only becomes embarrassing in the published document: snippets of conversation from the chat session, stage directions that break out of the argumentative structure, thank-yous to the language model, leftover placeholders, pushy LLM style patterns — and an AI disclosure that is missing, misplaced, or no longer true. This skill is the systematic cleaning pass before publishing: inspect, clean conservatively, correct the disclosure. It never changes the substance — it removes what is not part of the work.

Audit Checklist

Five finding classes, from clear-cut (fix directly) to sensitive (mark only):

1. Chat Residue and Stage Directions (clear-cut → remove/fix)

Sentences that belong to the CREATION of the text, not to the text itself: "As discussed, we keep this part in the paper because...", "Here is the revised section:", "I'd be glad to add...", leftover prompt fragments, meta-comments to the client/requester. Detection principle: The sentence falls out of the text and argument structure — it addresses a conversation situation instead of the reader. When deleting, check if a substantive core needs to be saved (transfer explanation into a footnote/body text).

2. Placeholders and Work-in-Progress Markers (clear-cut → resolve)

[TODO: …], [insert reference], XXX, <example here>, empty sections with headings, "(source?)". Resolve them or — if unresolvable — transfer them as open tasks into the project TODO and remove from the deliverable.

3. LLM Acknowledgments and Anthropomorphic Expressions (clear-cut → remove)

Thanking ChatGPT/Claude/Gemini & Co. does not belong in the acknowledgments section — tools are not thanked, their usage is declared in the AI disclosure. Likewise, remove anthropomorphic phrasing about the tool ("the AI kindly suggested").

4. AI Disclosure (verify → correct)

• Present? If the document was created with AI assistance and the venue/project requires or foresees a disclosure: does the section exist?
• Correct? Does it describe the actual usage (neither understated nor exaggerated)? Does it use the disclosure schema of the project/venue if defined (e.g., tiered levels)?
• Properly placed? At the venue-standard location (methods/acknowledgments area/dedicated section), identical across all language versions.

5. LLM Style Patterns (sensitive → fix clear cases only, mark the rest)

Formulaic transitions ("In summary, it can be said", "It is important to emphasize"), bullet-point inflation where running text belongs, "not only ... but also" chains, em-dash density, hedging phrases, and in English, well-known markers (including "delve", "tapestry", "it's worth noting"). Caution: Style is author territory — smooth out only unambiguous formulaic phrasing; present everything else as a findings list to the author instead of rewriting the text. A human-sounding text is not the goal of the skill; the goal is a text free of foreign bodies.

Workflow

1. Clarify scope: Which deliverables (files), which language versions? ALWAYS apply changes synchronously across all versions (cross-check: bilingual-doc-sync).
2. Mechanical scan: Full-text search for signal patterns (table below) — cheap, finds class 2/3 and parts of class 1 reliably.
3. Reading pass: Read the document along the argument structure — class 1 findings are recognised structurally (sentence addresses conversation instead of reader). Check especially: section beginnings/ends, acknowledgments, introduction/conclusion (residue lands there first).
4. Clean up: Fix classes 1–3 directly (conservatively, preserving substance), correct class 4, output class 5 as a findings list; smooth out directly only unequivocal cases.
5. Document: Record what was found/changed/marked — for papers with versioning obligations, note whether a new version/re-upload is necessary.
6. Periodic pass over a repository: Combine with rotation-check (one document/project per run, registry as memory).

Signal Patterns for Mechanical Scanning

The table is a starting point, not a filter substitute: patterns deliver candidates; the decision is made in context (steps 3–4). For purely mechanical character hygiene (emoji scan, control characters, broken umlauts), use existing tools — encoding damage is encoding-fix territory, not this skill's.

Example & Application

Request: "Check the paper for AI residue before upload."

1. Scope: paper_de.tex + paper_en.tex.
2. Scan: 1× "as discussed" (EN, section 4), 1× "[TODO: insert reference Smith]" (both),
   acknowledgments mention "valuable help from Claude".
3. Reading pass: In the introduction, a sentence directly addressing the reviewer
   ("We address this objection as requested in 3.2") → stage direction.
4. Fixes: Stage direction deleted (content was already in 3.2), TODO transferred as a task
   to TODO.md + placeholder removed, LLM thank-you deleted, instead
   AI disclosure section clarified regarding actual usage — all in DE and EN.
5. Note: Substantive change → new paper version required, entered into TODO.md.

Red Flags

Related Skills

• encoding-fix — Byte/encoding repair (mojibake); this skill works on the content level.
• bilingual-doc-sync — Keeping language versions in sync where fixes are applied.
• rotation-check — Scaffolding for periodic runs across a document repository.
• textproduction — Text generation (this skill is the QA afterwards).

Changelog

1.0.0 (2026-07-04)

• Initial version. Abstracted from Codex automation "research-llm-muster-check" (chat fragments in papers, LLM thank-yous, AI disclosure) and generalized to arbitrary deliverable texts; audit catalog extended with placeholders, style patterns, and scan signal table.