AI Slop Lint

A review-and-rewrite tool for textual artifacts (prose, docs, commit messages, PR descriptions, comments, READMEs). Detects the tonal and structural patterns that signal low-effort AI-generated or AI-assisted writing, explains why each one is a problem, and optionally produces a cleaned-up rewrite.

This skill is text-only. It does not analyze code structure, factual correctness, or authorship in a legal sense. Treat the output as a craft signal, not a verdict on who wrote what.

When to use this skill

Trigger any time the user wants prose evaluated or rewritten for AI-style tells. Examples:

"Does this README sound AI-generated?"
"De-slop this commit message."
"Clean up this draft — feels hedgy and generic."
"Review this blog intro for tone."
"Tighten this PR description."
User pastes a block of text and asks for a rewrite, tone pass, or review.
User points at a file and asks for a review with AI-slop concerns implied (e.g. "this doc reads like ChatGPT wrote it — help").

Do not trigger for:

Code review or code-quality questions (different skill territory).
Fiction or creative writing where the user wants an elevated or stylized voice.
Marketing copy the user has explicitly asked to be hype-heavy or punchy.
Factuality / hallucination checking — flag template placeholders and obvious refs, but do not verify claims.

Workflow

Identify the target text. It may be pasted, quoted, at a file path, or in the recent conversation context. If ambiguous, ask once which block the user wants reviewed.
Identify the user's intent. Three modes:
- Review only — default if unclear. Produce the findings report (see Report Format) without rewriting.
- Review + rewrite (minimal) — if the user says "rewrite," "clean up," "de-slop," "tighten," or similar. Preserve meaning, cut the slop. This is the usual ask.
- Review + rewrite (aggressive) — if the user says "aggressive," "rewrite hard," "cut ruthlessly," or similar. Allowed to delete whole sentences/paragraphs (vapid intros, chat sign-offs, disclaimers) rather than paraphrasing them.
Scan the text against the pattern cheat sheet below. For ambiguous or unfamiliar patterns, consult references/taxonomy.md for the full catalog.
Produce the report using the template in Report Format. Be specific — quote the offending span, name the pattern, say why it's slop, and suggest a fix.
If rewriting, produce the rewrite after the report as a clean markdown block, then (optionally) a brief "changes made" list for transparency. Present it as a preview — do not touch any files yet.
Ask before applying. After the rewrite block, explicitly ask the user whether they want you to apply it. See Applying the rewrite for the exact prompts by source type. Never overwrite a file, edit a paste target, or save anything without the user saying "yes, apply it" (or equivalent) in a separate turn.
Close with a handoff note (see Handoff note). This skill removes the AI-slop smell but does not know the user's house style, brand voice, commit-message convention, doc template, or target audience. Remind the user that the output is de-slopped prose, not yet styled prose — and suggest they run any tone/voice/format skills or workflows of their own on top of it before publishing.

Pattern cheat sheet

These are the patterns worth memorizing — they cover the majority of real-world cases. Full catalog with detection heuristics and examples lives in references/taxonomy.md — read that when a case feels borderline or you need the less common categories.

Very strong signals (near-definitive AI artifacts)

Flag on sight. These almost never belong in human-edited text.

AI self-references: "As a language model, I…", "As an AI, I cannot…", "As of my last update in 2023/2024/…", "my knowledge cutoff is…".
Prompt refusals / capability statements: "I cannot browse the internet", "I don't have access to real-time data", "I cannot directly edit…".
Chat sign-offs in static text (docs, commits, comments): "I hope this helps", "Let me know if you need anything else", "What would you like me to elaborate on?", "You're absolutely right!", "Great question!".
Markdown artifacts in non-Markdown contexts: raw **bold**, ### Heading, or backticks showing up inside a plain-text commit message, email body, or code comment where Markdown won't render.
Tool / citation placeholder traces: turn_0_search_0, {"source": "tool", "index": 0}, access-date=2025-xx-xx, bracketed citation scaffolds left unfilled.

Strong signals

Mid-sentence questions with canned answers: "The solution? It's simpler than you think." / "The real challenge? Staying ahead." — short question fragment followed by a glib answer. Very heavy tell in technical prose.
Unearned profundity beats: standalone dramatic sentences like "Something shifted.", "Everything changed.", "But here's the thing.", "And then it hit me." — especially when the surrounding text is neutral or technical.
Promotional tone in technical/neutral contexts: "groundbreaking," "revolutionary," "breathtaking," "dynamic atmosphere," "rich cultural heritage" showing up in a README, RFC, or commit message.
Elegant variation: cycling through near-synonyms to refer to the same thing ("the artist… the creator… the painter…"). Humans reuse the most accurate noun; models avoid repetition too aggressively.

Moderate signals

Usually worth flagging, but one alone isn't damning — the danger is clusters.

Vapid openers: "In today's fast-paced [landscape/world/ecosystem]…", "As technology continues to evolve…", "In an increasingly interconnected world…", "At the end of the day…".
Hedging / importance disclaimers: "It is important to note that…", "It's worth noting that…", "One must remember that…". Count them — ≥ 3 per 500 words is a smell.
Didactic framing: "In this section, we will explore…", "This article will delve into…", "Let's take a closer look at…" — especially in non-tutorial contexts.
AI vocab cluster: hits from this list appearing together in a short window (roughly 150–200 words, 3+ hits = strong cluster): delve, delving, tapestry, landscape, realm, ecosystem, intricacies, interplay, synergy, journey, navigate, harness, empower, foster, unlock, leverage, embrace, enhance, underscore, highlight, elevate, revolutionize, pivotal, crucial, transformative, profound, vibrant, dynamic, rich (when abstract), robust, scalable (when decorative).
Not-X-but-Y rhetorical parallelism: "It's not X. It's Y.", "This isn't just about X — it's about Y." Fine once; cumulative repetition is a tell.
Superficial analysis phrasing: "This underscores the importance of…", "serves as a powerful reminder that…", "stands as a testament to…", "marks a pivotal moment in the evolving landscape of…" — often attached to no concrete consequence.
Mismatched "from X to Y" ranges: "From the Big Bang to blockchain…", "From cell biology to dark energy…" — sweeping scope without real breadth.
Vague attributions: "Experts say…", "Observers have noted…", "Many believe that…" with no named source.
Em dash overload: em dashes used as catch-all punctuation, more than roughly once per 100 words in non-literary prose.
Excessive emoji bullets in non-marketing contexts (✅ 🚀 💡 as bullet points in a README or RFC).

Weak signals

Report only when they appear alongside stronger ones.

Rule-of-three triads ("fast, efficient, and reliable").
Formulaic conclusions ("In summary…", "In conclusion, this highlights the importance of…").
Monotonous sentence rhythm / repeated sentence starts ("However," "In addition," "This means" three or more times in a row).
Title-case headings where sentence-case would be standard.
Mixed curly and straight quotes.

deslop-ai-lint-skill