prose-sanitiser

name

prose-sanitiser

description

Audit and fix AI writing tells in prose, markdown, presentation content, and fiction. Removes Claude/LLM fingerprints at three levels: lexical (banned vocabulary, hedge words), structural (em-dash density, "The X" headings, negative parallelism, throat-clearing, transition overuse), and narrative (thematic over-explanation, embodied-emotion overuse, single-track plots, tidy resolutions, vague allusions, no fourth-wall breaks). Also provides generative principles for writing fresh prose that does not need sanitising. Enforces UK English spelling. Use when writing or editing public-facing content, presentations, documentation, articles, blog posts, tutorials, or short fiction. Invoke with: "sanitise this", "clean up the writing", "remove AI tells", "de-slop", "make this read human", "write a blog post", "draft technical content".

Prose Sanitiser

Strip LLM writing fingerprints from text. Output should read as if written by a competent human with opinions, not by a model hedging its way through a prompt.

UK English throughout. No exceptions.

Two modes:

• Generative — apply Section A principles when drafting new content.
• Destructive — run the Section B audit on existing text.

Mechanical pre-pass: the scanner

scripts/slop_scan.py catches the mechanical tells (Tier-1 vocab, "The X" headings, negative parallelism, throat-clearing, hedge words, US spelling, em-dash and transition density) before you read for the ones a regex cannot see. Run it first, fix in priority order, then do the human read for Section C and for voice.

python3 scripts/slop_scan.py <path>                 # full report + slop score
python3 scripts/slop_scan.py <path> --severity high # only the strongest signals
python3 scripts/slop_scan.py <path> --json          # machine-readable, for CI

It scans .md .markdown .mdx .txt .rst, skips fenced code and blockquotes, respects the slop-ignore marker (see below), and reports each finding with file:line and the fix. The exit code is the high-severity count, so CI can gate a docs build on it. The scanner sees lexical and structural tells only; it is blind to narrative defaults, altitude, and whether a sentence is actually true. Those are still yours.

Don't launder slop into a new slop (second-order defaults)

The failure mode of every de-slop pass is swapping one default for another. Kill every "leverage" and the prose acquires a different fingerprint: uniform "use", staccato two-word fragments ("Fast. Actually fast."), the same inverted "X, not Y" cadence on every other line, hedges amputated until the voice reads as clipped and machine-confident. An editor can clock a de-slopped-by-AI draft as fast as a slopped one. The replacement vocabulary, applied mechanically, is a tell.

So the rules below are a detector, not a target. The replace-with column is a prompt to make a choice, not a lookup table to apply on autopilot. The only durable property is the one a default can never have: a wording you chose for this sentence and can say why. Vary the repair. Sometimes "leverage" wants "use", sometimes "lean on", sometimes the clause should be cut. If a fix introduces a new uniform default, it is not a fix.

---

Section A — Generative Principles

Use when writing fresh content. Following these means less to fix later.

A1. Lead with value

The first 30 seconds decide whether the reader keeps going. Open with the specific thing they will learn, the concrete problem, or the result. Cut every warm-up sentence.

Bad	Good
"In today's rapidly evolving landscape..."	"Here's how we cut bug-detection time from 4 days to 2 hours."
"This article will explore..."	"We replaced our test runner. CI is 40% faster. Trade-offs below."

A2. Show, don't tell

Specifics beat adjectives. Numbers, names, code, before/after.

Bad	Good
"We improved testing."	"Bug detection: 12 → 47 per sprint."
"Performance improved."	"Response time: 2.3s → 180ms."
"Better collaboration."	"Devs now ask QE for input during story refinement."

A3. Honest trade-offs

Real writing names what is lost as well as gained. AI prose tends to claim "best of both worlds." Don't.

"TDD slowed velocity 20% in the first month. Bugs in production dropped 75% over the next quarter."

A4. Audience framing

Adjust the opening, the level of detail, and the takeaway:

• Developers — lead with the code or concrete problem; show implementation; discuss alternatives; link to repos.
• QA / QE — start with the testing challenge; show strategy not tools; include risk assessment; provide adaptable heuristics.
• Leadership — open with business impact; metrics that matter; connect technical decisions to outcomes; keep details concise.

A5. Write from experience

Only write about what you have done in production. If exploring, say so. The reader can tell when prose is generated from a vague middle distance rather than from concrete recall.

---

Section B — Destructive Audit

Work through every item. Fix in-place. Do not add explanatory comments.

B1. Em Dash (—) Density

Rule: Maximum 2 per 500 words in prose. Zero in lists.

Replace with:

• Comma (most cases)
• Full stop + new sentence (if the clause is independent)
• Colon (if introducing an explanation)
• Parentheses (if genuinely parenthetical)

Acceptable uses: attribution lines, dialogue interruption, deliberate rhetorical pause in a presentation heading.

B2. "The" Heading Disease

Rule: Never start a heading with "The" unless it's a proper noun ("The Guardian", "The Loop").

Before	After
The Problem: AI Is Hard	AI Is Hard
The Uncomfortable Question	One Awkward Question
The VisionFlow Stack	VisionFlow Stack
The 80/10 Gap	80/10 Gap

B3. Negative Parallelism

Rule: Kill "not X — Y" and "not X, but Y" constructions.

Before	After
It's not a feature — it's the foundation	It's the foundation, not a feature
Not just fast, but revolutionary	Fast. Actually fast.
This isn't about X. It's about Y.	This is about Y.

Invert: lead with the positive claim. Or just delete the negative half.

B4. Tier 1 Banned Vocabulary

Flag and replace every instance:

Kill	Replace with
delve	look at, examine, dig into
leverage	use
robust	solid, reliable, sturdy
seamless	smooth, clean
comprehensive	thorough, full, complete
cutting-edge	current, recent, new
transformative	(delete or be specific about what changed)
groundbreaking	new, first
innovative	(delete — show don't tell)
holistic	(delete or say "whole-system")
testament	proof, evidence, sign
tapestry	(delete — almost always slop)
vibrant	(be specific — what colour, what energy?)
utilize	use
harness	use
unlock	enable, open
unleash	release, enable
streamline	simplify, speed up
empower	enable, let, give
elevate	raise, improve
paradigm	model, approach
unprecedented	new, first, unusual
synergy	(delete)
optimize	improve, tune
foster	support, grow
underscore	show, highlight
navigate (figurative)	deal with, work through
ecosystem (when not biological)	system, network
deep dive	close look
game-changing	(delete or be specific)
enterprise-scale	production-grade, serious, large
enterprise-grade	production-grade
extraordinary	(delete or be specific)

B5. Tier 2 Cluster Words

Not banned individually, but flag if 3+ appear in a single section:

crucial, notable, noteworthy, remarkable, fascinating, profound, compelling, intriguing, elegant, meticulous, intricate, deliberate, thoughtful, sophisticated, sprawling, bustling, evocative, poignant, cornerstone, linchpin, bedrock, nexus, interplay, realm, arena, sphere, endeavour, myriad, plethora

B6. Throat-Clearing Openers

Delete entirely:

• "In today's rapidly evolving..."
• "In the world of..."
• "Here's the thing:"
• "Let me be clear:"
• "It turns out..."
• "Let's dive in / explore / unpack"
• "It's worth noting that..."
• "It's important to note that..."
• "At its core..."
• "At the end of the day..."
• "When it comes to..."

B7. Sycophantic Filler

Delete entirely:

• "You're absolutely right"
• "Great question"
• "That's a really interesting point"
• "Certainly!"
• "Absolutely!"
• "I'd be happy to help"

B8. Hedge Words

Flag and usually cut:

basically, actually, probably, essentially, fundamentally, very, really, quite, perhaps, somewhat

If a claim needs a hedge, replace with a specific qualifier ("in the staging environment", "for payloads under 10KB").

B9. Structural Tells

Check and fix:

• Rule of three: If you have exactly 3 items in every list, vary it. Real writing has 2s, 4s, and 7s.
• Uniform paragraph length: Vary it. Short paragraphs hit harder.
• Tell-show-summarise: Delete the summary sentence at the end of each section. The reader just read it.
• Stacked rhetorical questions: Maximum 2 in sequence. Better: 1 question, then answer it.
• Bold-label bullets: Not every bullet needs a Bold Term: prefix. Reserve for reference material.
• Copula substitution: "serves as a" → "is". "marks the" → "is".

B10. Transition Word Overuse

Flag if more than 2 per page:

Furthermore, moreover, additionally, consequently, notably, crucially, importantly, ultimately, fundamentally, indeed, significantly, subsequently, accordingly

Replace with: nothing (just start the next sentence), or a concrete connector that adds information.

B11. Passive Voice

Rule: Active by default. Passive only when the actor is genuinely unknown or irrelevant.

Before	After
It can be seen that...	This shows...
The decision was made to...	We decided to...
The system is designed to...	The system does...

B12. UK English Spelling

Enforce throughout:

US	UK
optimize	optimise
organize	organise
color	colour
behavior	behaviour
center	centre
license (noun)	licence
defense	defence
analyze	analyse
catalog	catalogue
fulfill	fulfil

---

Section C — Narrative Tells (Fiction)

Apply when sanitising short stories, novel passages, character-driven scenes, or any narrative prose. Lexical fixes are not enough: AI fiction converges on a narrow set of structural defaults. Source: Russell et al., StoryScope (arXiv:2604.03136, 2026), parallel corpus of 61,608 stories across 5 LLMs and human authors.

The percentages below are AI-vs-human rates from that study; treat them as a strong prior, not a hard rule.

C1. Thematic Over-Explanation

AI states the theme/moral explicitly 77% of the time vs. 52% for humans. Narrators announce the lesson learned. Subplots tidily echo the main theme.

Fix: Delete the narrator's thematic commentary. Let the reader infer. Allow subplots to drift. If a character's arc ends with stated insight, either cut the statement or move it into ambiguous action.

Slop	Repair
"She realised, then, that grief was simply love with nowhere to land."	(delete the line — show the realisation through what she does next)
Narrator: "And so, in the end, he learned that..."	(cut entirely)

C2. Embodied Emotion Over Labels

AI conveys emotion through body 81% of the time vs. 38% for humans. Tight throats, cold sweat, dimming light. Humans use explicit emotion labels 29% of the time; AI only 8%.

Fix: Mix. Sometimes name the feeling ("she was furious"). Use bodily sensation sparingly and only where it earns its place. Stop using setting as mood mirror in every scene.

Slop	Repair
"A cold weight settled in his chest. The lamplight dimmed. The wallpaper seemed to lean in."	"He was scared." (or: a single concrete bodily detail, then move)

C3. Single-Track Plots

79% of AI stories have no subplots, vs. 57% for humans. Causal chains are tight; loose ends are tied. Protagonists drive their own resolution 69% of the time (vs. 46% human).

Fix: Introduce at least one subplot that does not resolve. Let an event have multiple causes, some off-page. Let the resolution come partly from chance, secondary characters, or refusal to act.

C4. Tidy Resolutions, Especially Internal-Acceptance Endings

AI defaults to "the protagonist understands and accepts" (47% vs. 27%). Epilogues are over-represented (a Claude fingerprint). Endings rarely unsettle.

Fix: Cut the epilogue. End on the ambiguous beat. Let the protagonist be wrong, or unchanged, or worse off. The reader does not need closure.

C5. Linear Chronology

AI tells the story from first clue to grand reveal. Humans use flashbacks, flash-forwards, and nonlinear framing far more often (anachrony intensity 2.58 vs. 2.31; nonlinear framing 1.96 vs. 1.68 on 1-5 scales).

Fix: Consider opening at the funeral and spiralling backward. Withhold the inciting incident. Let a revelation force re-reading of earlier scenes.

C6. Vague Allusions Over Named References

AI uses unnamed "implicit echoes" 72% of the time (vs. 50% human). It avoids naming real brands, places, works, or people. Humans cite specific texts and authors at nearly double the rate (47% vs. 24%).

Fix: Name the band, the pub, the novel, the brand of cigarette. Specificity signals a human who lived in a world rather than a model trained on its silhouette.

C7. No Fourth-Wall Breaks

Humans break the fourth wall 67% of the time vs. 39% for AI; they address the reader directly 28% vs. 7%.

Fix: If the voice permits it, let the narrator acknowledge the reader. "You may think this was foolish. It was." Use sparingly.

C8. Over-Engineered Sensory Description

AI over-indexes on olfactory imagery (82% vs. 57%) and lush sensory density generally. Spatial granularity runs higher than human writing.

Fix: Drop most of the smell descriptions. Leave rooms partially undrawn. Trust the reader's imagination to fill in walls, weather, light.

C9. Philosophical-Debate Dialogue

AI characters debate ideas in dialogue 59% of the time (vs. 34% human). Conversations become essays in disguise.

Fix: Cut philosophical exchanges. Replace with action, deflection, or talking past each other. People rarely argue ideas cleanly; they argue about who left the milk out.

C10. Morally Clear Protagonists

Human stories have morally ambivalent protagonists 59% of the time; AI only 38%. AI heroes do the right thing for the right reasons.

Fix: Give the protagonist a petty motive alongside the noble one. Let them be cruel in a small way. Decline to signal which choices are admirable.

C11. Per-Model Fingerprints

If you know which model generated the draft, watch for:

Model	Tell
Claude	Flat event escalation; reverent toward literary tradition; epilogues; avoids dream sequences; "quiet endings".
GPT	Dream sequences; gossip/rumour as plot mechanism (64% vs. 44-55%); distant retrospective framing ("years later..."); ensemble casts.
Gemini	External character description as the introduction; bleak/oppressive settings (88%); tidy denouements; extended endings.
DeepSeek	Front-loaded context that humans would withhold.
Kimi	The generic centre — few distinctive choices, blandly competent.

---

Section D — Final Editing Checklist

Before publishing:

• Title promises something specific
• Opening hooks in 30 seconds (no warm-up)
• Every claim is backed by a specific example, number, or quote
• All Tier 1 vocabulary removed
• Em-dash count under threshold; no em-dashes in lists
• No "The X" headings (unless proper noun)
• No negative parallelism
• No throat-clearing openers
• UK English consistent throughout
• (Fiction) at least one subplot doesn't tidily resolve
• (Fiction) at least one emotion labelled directly
• (Fiction) at least one named real-world reference
• Would send to a respected colleague without an apology

---

Output Format

Default: return the cleaned text with no commentary.

If asked for a report (or running an audit), lead with the verdict, not the findings:

1. Verdict and the single highest-impact change. One line: the slop score and verdict from the scanner, then the one fix that matters most.
2. Findings by priority, each with file:line, the tell, and the fix. High-severity first. Quote the offending span, not the whole paragraph.
3. Close with the slop score and the top three changes. Plain and specific.

The goal is prose a person decided the wording of, which is the one thing the scanner cannot do for them. State what is mechanical (the scanner found it) and what needed a human read (Section C, voice, truth).

When NOT to sanitise

• Direct quotes from other people (blockquotes — the scanner already skips these)
• Code, terminal output, API responses (the scanner skips fenced blocks)
• Proper nouns and product names
• Technical terms of art (even if they overlap with the banned list)
• Content the user explicitly flags as intentional
• Stylistic choices that violate a rule but serve the piece (e.g. a deliberately philosophical dialogue in a Socratic essay)

Marking intentional choices. Put slop-ignore on a line (an HTML comment <!-- slop-ignore --> works in markdown) and the scanner skips it. Use it when a flagged word is a real decision, so the audit stays trustworthy and does not nag about a chosen term. A line you have to mark is a line you have made a choice about — which is the whole point.