| name | ai-slop-detector |
| description | Universal prose audit. Scores writing on TWO axes — AI-Slop (does this read like AI wrote it?) and Comprehension (can a fresh reader follow this?). Use PROACTIVELY as a mandatory final QA pass on ANY prose generated for humans to read — every email (internal or external), proposal, report, status update, blog post, social post, doc, or chat message — before delivering it, even when nobody asked for an audit. Also use whenever the user wants to audit, critique, score, rewrite, or humanize prose: "audit this", "review this", "is this AI", "does this sound like AI", "humanize this", "make this less AI", "rewrite this", "de-slop this", "clean this up", "slop check", "score this", "is this readable", "would a fresh reader follow this", "comprehension check". Also the final pre-delivery pass inside other writing skills (cold-email, copywriting, sales-enablement, ad-creative, email-sequence, mahmouds-seo-writer, mahmouds-writing-voice). |
slop-cop
A universal prose audit with two parallel axes. Built on ~135 published sources spanning peer-reviewed linguistics, AI-detector vendor methodology, plain-language style guides (plainlanguage.gov, GOV.UK, Microsoft, Google), cognitive-load research (Miller, Sweller, Pinker), web-readability research (NN/g), and the Plain Writing Act / WCAG accessibility standards.
The one-line summary
Two axes, two verdicts. AI-Slop and Comprehension. A piece can pass one and fail the other.
- AI-Slop axis — Does this read like AI wrote it? Patterns, vocabulary, formatting, rhythm.
- Comprehension axis — Can a fresh reader follow this? Acronyms, named-entity bombing, telegraphic compression, readability, structure.
Single instances aren't a signal. Density is. Both axes use density-based scoring (per 500 words, weighted by severity) with the same verdict tiers (PASS / LOW / MEDIUM / HIGH / CRITICAL). The audit reports both and combines them into a single recommendation based on whichever is worse.
Why this skill exists
Our feeds are becoming shit. Our websites are becoming shit. Our repos are becoming shit. AI didn't make writing harder, it made writing easier, and now everyone uses the same shortcuts, the same shapes, the same words. Open ten landing pages in a row and you can't tell them apart.
I built this skill to make my own websites less shit, my clients' websites less shit, and my LinkedIn feed less of a copy-paste graveyard. Yes, AI wrote parts of this skill. That is not the problem. The problem is AI prose nobody catches: the safe, hedge-stacked, em-dash-heavy paragraph where every line is grammatically clean but the whole thing is forgettable.
I run this against my own work every day. Landing pages, blog posts, READMEs, pitch decks, cold emails. It catches the stuff I would have shipped. If someone uses it to score other people's writing, fine. The first job is policing yourself before you ship.
Running it daily means finding new things every week. A pattern I missed. A false alarm on real human prose. A new model with new tells. The list moves with the work. Expect a lot of changes.
Mode selection
Three modes. Audit is the default and changes nothing. The two rewrite modes are
opt-in. Pick one based on what the user wants.
| Signal | Mode |
|---|
| "audit this", "review", "critique", "is this AI", "is this readable", "score this", "check for slop" | Audit (default) |
| Another writing skill is wrapping up and about to deliver prose | Audit pass before delivery |
| "rewrite", "humanize", "make this less AI", "de-slop", "clean up", "polish" (text pasted in) | Rewrite |
| "fix this file", "edit this in place", "clean up this doc" (a file to modify) | Edit-in-place |
Audit scores and reports; it does not touch the text. Rewrite returns a cleaned
version plus a before/after score, leaving the original intact. Edit-in-place
applies targeted edits to a file and preserves passages that already read human.
Both rewrite modes follow the rewrite workflow below and use references/rewrite.md.
If ambiguous, ask one short question: "Audit only, or do you want a rewritten version?"
When to skip
This skill governs prose meant for human readers. Skip it for:
- Code, code comments, commit messages, or PR descriptions
- Technical API documentation or reference material (different audience target — see calibration §10)
- Raw data, structured outputs (JSON, YAML, CSV)
- Direct quotations from other people that should preserve their voice
- Instructions to other agents or skills (system prompts, agent briefs)
- Intentionally formal or legal documents (different audience target)
- Dialogue under another character's voice in fiction (apply voice-aware judgment)
The audit workflow
Five steps. The scanner does the mechanical work on both axes; reading does the qualitative work; calibration converts findings into two verdicts plus a combined recommendation.
Step 1 — Run the scanner
python3 scripts/scan.py /path/to/draft.md
echo "draft text..." | python3 scripts/scan.py
Flags:
python3 scripts/scan.py --quick draft.md
python3 scripts/scan.py --json draft.md
python3 scripts/scan.py --audience marketing draft.md
python3 scripts/scan.py --audience academic draft.md
python3 scripts/scan.py --audience technical draft.md
python3 scripts/scan.py --genre encyclopedic draft.md
python3 scripts/scan.py --strict-em-dash draft.md
The scanner outputs:
- Two verdicts (AI-Slop, Comprehension) with density scores
- Combined recommendation
- Per-axis violation breakdown
- Readability metrics panel (Flesch RE, FK Grade, SMOG, Coleman-Liau, Dale-Chall, lexical density, avg sentence length, passive voice %)
- Density signals (acronym density, named-entity density, numeric density per sentence)
- Burstiness, contraction ratio, model fingerprint
Step 2 — Read against both pattern catalogs
AI-Slop axis: Load references/patterns.md. Walk through the 47 patterns by group. The scanner catches the mechanically-detectable subset; the rest requires reading.
Comprehension axis: Load references/comprehension.md. Walk through the 35 patterns by group. Roughly 17 are mechanically detectable; the rest require reading. Particularly:
- Buried lede / missing thesis — does the first paragraph tell the reader the point?
- No topic sentences — does each paragraph open with its claim?
- Curse of knowledge — does the writer assume context the reader lacks?
- No concrete examples — does every abstract claim have a specific instance?
- Definition by synonym — are domain terms defined with concrete examples or just other jargon?
For each pattern (both axes), flag with quote + severity.
Step 3 — Apply calibration
Load references/calibration.md. Apply:
- Density-based scoring on both axes (§1 for AI-Slop, §9 for Comprehension)
- Genre adjustment (AI-Slop: §3) and audience calibration (Comprehension: §10)
- Compound triggers — escalate when 3+ H tells in one paragraph (slop) or any 100w window has 3+ undefined acronyms / 5+ named entities (comp)
- Cross-axis recommendation (§11) — single sentence based on whichever axis is worse
The scanner does most of this automatically. The reader applies judgment to ambiguous cases.
Step 4 — Output the audit report
Use the dual-verdict format in references/audit-report-template.md. The report has:
- Two verdicts (AI-Slop and Comprehension) in a header table
- Combined recommendation — one sentence drawn from cross-axis matrix
- Stats block — word count, audience, burstiness, model fingerprint
- AI-Slop axis section — counts, top fixes, mechanical violations, qualitative violations, calibration notes
- Comprehension axis section — counts, top fixes, readability metrics panel, density signals, mechanical violations, qualitative violations, calibration notes
- Combined top 3 fixes — pulled from both axes by impact
- Combined recommended action — what to do next
Don't soften findings on either axis. The point of the audit is to catch what the writer missed.
Step 5 — If asked, deliver the revision
If the user wanted a rewrite, not just critique, switch to the rewrite workflow
below. Do not free-hand a rewrite; the workflow measures the result so you can
prove it improved.
The rewrite workflow
Load references/rewrite.md for the full doctrine and references/voice-profiles.md
for the voices. The loop, in short:
1. Scan for a worklist. Run --suggest to get before-scores, deterministic
safe fixes (with offsets), and a judgment worklist with per-item guidance:
python3 scripts/scan.py --suggest path/to/draft.md
2. Apply the safe fixes. These are meaning-preserving 1:1 swaps only (em dash →
comma, "utilize" → "use", "in order to" → "to"). They skip text inside quotes and
code, so a word quoted AS an example is never rewritten. The engine owns these:
python3 scripts/scan.py --apply-safe path/to/draft.md
python3 scripts/scan.py --apply-safe --in-place path/to/draft.md
3. Rewrite — patch a little, rebuild a lot. Pick a voice profile. Default to the
warmest voice that fits the genre; "professional" reads as corporate unless the
context truly needs distance. The worklist tells you WHAT is wrong; how you fix it
depends on how much is wrong:
- Light draft (LOW/MEDIUM, scattered flags): patch in place — split long
sentences, name vague authorities, kill copula-avoidance, define acronyms.
- Heavy draft (HIGH/CRITICAL or templated marketing/AI copy): rebuild it. Do NOT
edit span by span. Read the worklist as diagnosis, set the draft aside, and write
the piece fresh from its brief in the target voice. Change the skeleton: dissolve
bullet lists into prose, open on a concrete scene, reorder. Patching a heavy draft
keeps the corporate shape, and the shape is half the slop. Rebuild, don't renovate.
Either way, rewrite for both axes at once, use concreteness — replace a cut
abstraction with a specific (a scene, a name, one true detail), never a warmer
abstraction — and cut to the bone: a rebuild should come out shorter than the
slop it replaces, not longer. Stay evidence-safe: keep every element of the brief,
change no facts or numbers, invent nothing. See references/rewrite.md.
4. Check fidelity, then re-scan and prove it. First, fidelity: list every claim,
feature, audience, and offer in the original and confirm each survived the rewrite,
and that you invented nothing new. Getting concrete by dropping an element or adding
a fake detail is a fidelity failure, not a rewrite. Then run the plain scan and report
before/after verdict and density on both axes. Check burstiness too: if it sits
below about 0.5, or the scan reports sanded_prose, the rewrite went sterile — vary
the rhythm and add a concrete detail, don't ship it flat. If either axis is still
above LOW, do one more pass on what remains. Two passes is normal; a third means the
draft needed a rethink.
5. Deliver. For rewrite mode, return the cleaned text plus the before/after
scores. For edit-in-place, apply the changes with the Edit tool, preserving
passages that already scanned clean, and report what changed. The re-scan is the
proof; never claim a rewrite worked without it.
Quick reference: lethal tells (both axes)
If you only have time for a quick scan, look for these. Each appears in the highest-density failures.
AI-Slop (most lethal)
Vocabulary: delve / delves, tapestry, underscore / underscores, leverage (verb), harness
Sentence-level: "It's not X, it's Y" (negation reversal), "serves as a" / "stands as a" / "boasts" (copula avoidance), "X happened, demonstrating Y" (-ing tail), "From small startups to global enterprises" (false range), "Studies show..." with no citation
Voice: "Great question!" (opener sycophancy), "I hope this helps!" (closer sycophancy), "In today's fast-paced world..." (107x more in AI), "As a society, we must..." (royal we), "As of my last update..." (knowledge-cutoff leakage)
Structural: "In conclusion / Furthermore / Moreover" (listicle transitions), "It's worth noting that..." (throat-clearing), em dashes in clusters (3+ per 500w), bold-first bullets, "X: A Comprehensive Guide" titles
Abstraction & agency: "The stakes are high" / "the implications are significant" (vague declaratives — gravity with no specific thing), "Here's the thing" / "the truth is" / "let that sink in" (throat-clearing openers + emphasis crutches), "the data tells us" / "speaks for itself" (false agency — name the human)
Comprehension (10 most lethal)
- Undefined acronyms — 3+ per 100 words = high cognitive load
- Named-entity bombing — 5+ unfamiliar proper nouns per 100 words
- Stat bombing — 3+ numerics in one sentence with no comparative anchor
- Telegraphic colon-labeling — Anchor case: X. Tools that win: Y. Compresses topics into label-list rather than prose
- Coined insider terms — "two-stack social management thesis" used as if known
- Long sentences — over 30 words; comprehension drops sharply
- Buried lede — main point arrives after 2+ paragraphs of setup
- Missing thesis — reader can't summarize the central claim after reading
- Wall-of-text — paragraphs over 100 words with no breathing room
- No concrete examples — every abstract claim left abstract
If a draft has 5+ AI-Slop lethal items in 500 words → AI-Slop HIGH/CRITICAL.
If a draft has 3+ Comprehension lethal items in 500 words → Comprehension HIGH/CRITICAL.
Calibration principles
A short version of the rules in references/calibration.md. Read the full file before producing the calibration section of the audit.
Density formula (both axes)
density = (H × 3) + (M × 1) + (L × 0.25) per 500 words
Verdict thresholds (both axes)
- 0–2 = PASS
- 2–5 = LOW
- 5–10 = MEDIUM
- 10–18 = HIGH
- 18+ = CRITICAL
AI-Slop axis tuning
- Genre adjustments: academic prose can use "studies show" with citations; marketing tolerates more intensifiers; encyclopedic prose triggers false positives (LLMs were trained on Wikipedia); fiction respects character voice
- Contested tells: em dashes in clusters = H, alone = L (Mahmoud-mode = always H via
--strict-em-dash); "actually" survives only contrasting reality with theory
- Sanded-prose alert: if famous vocabulary is clean but structural tells are heavy, flag — writer prompt-engineered around the v1 list
- Uncanny valley: if 8+ weak tells stack with burstiness below 0.5, escalate one tier
Comprehension axis tuning
- Audience calibration (calibration.md §10): the SAME prose hits different verdicts depending on audience. General web/blog targets FK grade 7–9; marketing 6–8; technical 10–12; academic 12–16. The scanner uses
--audience to pick the threshold band.
- Compound triggers escalate one tier:
- 3+ undefined acronyms in any 100-word window
- 5+ named entities in any 100-word window
- 3+ telegraphic colon-labels in one paragraph
- Any paragraph over 150 words with no subheading
- Any sentence over 40 words
- Readability metrics calibrate, patterns rule: the metrics panel (FK Grade, lexical density, etc.) is diagnostic. The verdict comes from the catalog patterns. When metrics and patterns disagree, call it out (e.g., "patterns clean but FK Grade 16 — academic-density texture without specific failures").
Cross-axis recommendation
| AI-Slop | Comprehension | Recommendation |
|---|
| PASS / LOW | PASS / LOW | Ship it. |
| MEDIUM+ | PASS / LOW | AI-Slop revision; reader-friendly already. |
| PASS / LOW | MEDIUM+ | Comprehension revision; texture is fine. |
| MEDIUM+ | MEDIUM+ | Both cleanup; often overlapping fixes. |
| HIGH+ | HIGH+ | Full rewrite. |
Reference loading guide
Read what the task needs.
| File | Read when |
|---|
references/patterns.md | AI-Slop qualitative pass. The 47 rhetorical/structural patterns. |
references/vocabulary.md | AI-Slop word-level tells. 200+ items in tiered categories. |
references/formatting-tells.md | AI-Slop formatting checks. 33 items in 5 categories. |
references/comprehension.md | Comprehension qualitative pass. The 35 patterns in 5 groups. |
references/readability-metrics.md | When interpreting the readability panel. The 8 formulas + thresholds + audience targets. |
references/rewrite.md | Both rewrite modes. The rewrite loop, evidence-safe rules, patch-vs-rewrite threshold. |
references/voice-profiles.md | During a rewrite. Seven voices + which --audience each pairs with. |
references/calibration.md | Always at the start of every audit. Density rules, genre, audience, fingerprints, contested tells, cross-axis matrix. |
references/audit-report-template.md | Always at the start. Defines the dual-verdict output format. |
references/sources.md | When challenged on a specific tell. ~130 sources backing the catalogs. |
The scripts/scan.py scanner runs both axes in one pass and produces a report that references the categories and items in these files.
A note on the detector itself
This is a shape-and-comprehension detector, not a classifier. The AI-Slop axis measures whether prose has the shape of AI writing — patterns, vocabulary, structure, rhythm. The Comprehension axis measures whether a fresh reader can follow it. Neither tells you whether AI wrote a given piece.
A skilled writer using AI as a research tool can produce prose that scores PASS on both axes. A careful prompt can produce LOW on both. A human writing fast can produce HIGH on Comprehension while PASSing AI-Slop. A polished AI marketing post can score CRITICAL on AI-Slop while readable.
Treat the AI-Slop verdict as "this prose has the shape of AI writing." Treat the Comprehension verdict as "a cold reader will struggle here." Both verdicts are actionable.
The catalog will need updating as models change. After "delve" went viral in early 2024, arXiv frequency dropped sharply. After GPT-5.1 added an em-dash opt-out (Nov 2025), em-dash density became less reliable. The skill stays useful by weighting newer patterns higher and surfacing density rather than individual hits.
When the scanner says PASS but the prose still reads wrong, trust the reading — the qualitative patterns (parallel structure, force of metaphors, voice consistency, curse-of-knowledge moments) are what humans pick up first and what regex misses last.
