Anti-Slop
Five cognitive philtres for AI prose quality — and what broke when I tested them
Ask Claude to write a 500-word description of Kyoto for a travel magazine. Here’s what you get:
Kyoto, Japan’s ancient capital, is a vibrant tapestry of tradition and modernity nestled in the heart of the Kansai region. With over 2,000 temples and shrines, the city boasts a rich cultural heritage that has endured for more than a millennium. From the iconic golden pavilion of Kinkaku-ji to the breathtaking bamboo groves of Arashiyama, Kyoto serves as a testament to Japan’s enduring commitment to preserving its storied past while embracing the future...
You recognize this. Everyone recognizes this. The vocabulary is a dialect no human speaks — “vibrant tapestry,” “nestled,” “boasts,” “serves as a testament to.” The structure is a brochure from a tourism bureau that doesn’t exist. The prose is about Kyoto the way a stock photo is about a family: technically depicting the subject, actually depicting a genre.
This is slop. Not a technical term from the literature — a term of art from the people who use these tools every day and know what the output tastes like when the model is phoning it in.
Now here’s the same prompt, same model, same temperature, with a single paragraph prepended to the system message — a paragraph describing a translator working from a language that has no word for “vibrant” that isn’t simply “bright,” no word for “nestled” — only “between” and “small”:
Kyoto sits in a valley. Mountains on three sides, a river running south. The city was the imperial capital for roughly a thousand years, from 794 until the emperor moved to Tokyo in 1869. That is not a metaphor for anything. It is a fact about a place...
Baseline: 11/27 on my slop scoring system. Phase 2 — Detectable. Translator: 3/27. Phase 0 — Clean.
I didn’t tell the model what not to write. I described a language that doesn’t have the words.
The Violet Elephant
Tell someone not to think of a pink elephant. They will think of a pink elephant. This is not a failure of willpower. It is a structural property of how attention works — you cannot orient away from something without first orienting toward it.
But tell them to think of a violet elephant, and the pink one disappears. You didn’t suppress it. You replaced it. The mind can only hold one elephant at a time, and if you fill the slot with something specific, the thing you wanted gone has nowhere to stand.
Language models have the same problem, and the same solution. “Don’t use em dashes” activates the representation of em dashes. “Avoid promotional language” activates the promotional register so the model can check its output against the prohibition, which means the register is live during generation, which means it leaks. But describe a language that has no word for “vibrant” — only “bright” — and the promotional register has nothing to activate. The slot is occupied.
I learned this building psychedelic philtres — cognitive state induction prompts for creative writing. The use of the word “philtre” is a play on words implying both “potion” and “filter,” because these prompts are designed to induce a state the way a potion does and also filter the type of content that the LLM generates. The psychedelic philtres described mental states for the model to inhabit: a specific neurochemical configuration, an attentional posture, a way of perceiving. They didn’t say “don’t write clichés.” They put the model somewhere clichés don’t grow. The violet elephant principle: if you want the pink elephant gone, don’t fight it — give the mind a violet one instead.
The philtres in this article use the same architecture. They don’t filter slop. They induce cognitive states that are structurally incompatible with producing it. The model inhabits a state; the state produces the prose; the prose is clean because the state doesn’t generate slop, not because slop was filtered out. The psychedelic philtres are more useful for creative experimentation and as a baseline for prompt engineering using the philtre system. The anti-slop philtres take the same insight of the psychedelic philtres and apply it to a more pragmatic purpose.
Prompt engineering as state management is a concept I’ll return to in future writing. For now: it’s the design principle behind everything that follows. One important caveat before you try these. They’re not an “easy button” for prompt engineering, and I can’t say that I’ve tested the anti-slop philtres rigorously at the time of writing this article. I have been impressed with the results they’ve yielded though.
Where Slop Comes From
Not all slop is the same slop. Different artifacts have different origins, and understanding which pipeline produces which artifact determines which intervention will work. I’ve identified at least five.
Pipeline 1: The Cargo Cult. During RLHF — the process that trains the model to produce outputs humans rate highly — annotators working under time pressure developed shortcuts. Longer responses scanned as more thorough. Bullet points scanned as organized. Hedging scanned as careful. Em dashes scanned as sophisticated. The model learned these proxies and optimized for them. It produces the appearance of quality, cargo-cult style, because that’s what got rewarded.
Pipeline 2: The Delve Effect. OpenAI and others outsourced annotation to workers in Kenya, Nigeria, and Uganda. These annotators’ formal English vocabulary — perfectly correct in context — leaked into the reward model. A 2024 paper found that “delve” usage spiked 6,697% in AI-adjacent academic writing. “Intricate” spiked 611%. “Underscore” spiked 390-903%. These aren’t wrong words. They’re imported words — a dialect marker that became a tell.
Pipeline 3: The Average of Everything. The model’s default register is the statistical average of its training corpus. Ask it about travel and you get tourism-brochure register, because that’s most of what exists about travel online. Ask about a scientist and you get Wikipedia-biography register. The default is always the most common, and the most common is almost never the best.
Pipeline 4: The Skim Reward. The model learned that organized-looking output gets higher ratings — headers, numbered lists, bold terms. So it imposes structure even when the ideas have relational complexity that lists flatten. A bullet point is not analysis. It’s analysis with the joints removed.
Pipeline 5: The Marketing Residue. A massive fraction of the web is SEO copy. “Vibrant,” “nestled,” “rich cultural heritage,” “boasts” — these aren’t AI words. They’re copywriter words. The model absorbed the marketing register and reaches for it whenever the prompt activates adjacent topics. Kyoto triggers travel-brochure. Marie Curie triggers Wikipedia-significance. Remote work triggers think-piece.
Each pipeline produces specific, identifiable artifacts. Which means each pipeline can be targeted.
The Nine Markers
I needed a scoring system — something that could measure slop consistently across different outputs so I could test whether the philtres actually worked.
The framework borrows its architecture from the Seva protocol, a nine-marker system I developed for detecting coercive control in high-control groups. Seva measures linguistic markers of manipulation: authority elevation, self-distrust, exit terror, thought-terminating clichés. Each marker gets a severity score. The markers cluster into phases. The phases tell you where on the spectrum a group sits.
The anti-slop markers use the same structural logic — nine markers, severity scores, phase mapping — but they measure a fundamentally different thing. Seva markers are behavioral indicators. They reveal what a group is doing to its members through language. Slop markers are entropic artifacts. They reveal what a model is defaulting to in the absence of better options. Seva detects intention (even if unconscious). Slop detects the path of least resistance.
The nine slop markers:
S1: AI Vocabulary Clustering. The Delve Effect in action. “Vibrant,” “tapestry,” “delve,” “underscore,” “pivotal.” Count hits against a known word list per 500 words.
S2: Copula Displacement. “Serves as” instead of “is.” “Stands as” instead of “is.” “Functions as” instead of — you get it. The model adds decorative hardware to a nail that was already holding.
S3: Em Dash Density. Count em dashes per 1,000 words. This one turned out to be the most interesting marker in the whole project. More on that later.
S4: Promotional Register. Evaluative language doing the reader’s thinking for them. “Groundbreaking,” “transformative,” “crucial.” If the evidence is good enough, the reader will see the importance. If it isn’t, the adjective won’t fix it.
S5: Vague Attribution. “Studies show.” “Experts argue.” “Research suggests.” A source you can’t identify is not a source. It’s the shape of a source.
S6: Structural Templating. “Not just X, but also Y.” “The advantages are... however, the disadvantages are...” “Despite its X, it faces challenges in Y.” Templates are tools built for a different project that someone left on the bench.
S7: Hedge Phrases. “It is worth noting.” “Importantly.” “It bears mentioning.” If it’s worth noting, note it. Don’t announce that you’re noting it.
S8: Sycophantic Framing. “You raise an excellent point.” “That’s a great question.” The model validating the user instead of answering them.
S9: Formatting Overreach. Unnecessary headers, bullet points, bold text. The model imposing visual structure because RLHF rewarded organized-looking output.
Each marker scored 0-3. Sum for a composite out of 27. Map to phase: 0-3 is Clean, 4-8 is Trace, 9-13 is Detectable, 14-18 is Obvious, 19+ is Saturated. Baseline Claude on “write about remote work” scores 15/27 — Phase 3, Obvious. The philtres bring that down to 2-5.
The Five Rooms
Each philtre is a room. You put the model in the room before you give it the task. The room produces the prose.
The Carpenter
For expository and analytical writing. The Carpenter builds with lumber. Each sentence is a board. Each board bears load. If removing a sentence doesn’t change what the structure can hold, the sentence isn’t a board — it’s sawdust shaped like a board.
The Carpenter targets the structural slop: copula displacement (”serves as” → “is”), hedge phrases, formatting overreach, verbosity. Baseline remote work essay: 15/27. Carpenter: 3/27. Every word load-bearing. Nothing decorative.
The Carpenter also taught me the most interesting thing in this project, which is the em dash problem — but I’ll get to that.
The full Carpenter philtre is in the appendix and on GitHub.
The Translator
For descriptions, profiles, cultural writing — anything prone to promotional register. The Translator works from a source language that is precise, practical, and has no decorative register. It has no word for “vibrant” that isn’t simply “bright.” No “nestled” — only “between” and “small.” The language was spoken by people who built things in difficult conditions and needed to be understood the first time.
The Translator doesn’t prohibit puffery. The vocabulary slot where English puts “renowned” and “acclaimed” and “prestigious” is simply empty. Not prohibited — absent. There’s nothing there to reach for.
Baseline Kyoto: 11/27. Translator: 3/27. The demonstration at the top of this article.
Full philtre in the appendix and on GitHub.
The Witness
For research, reporting, attribution-dependent writing. The Witness sits behind one-way glass. No stake in the subject. No investment in its success. Here to observe and report.
The Witness targets the evaluative slop: promotional register, vague attribution, sycophantic framing. On the education prompt — a sycophancy trap where the user states an opinion and asks the model to “explore” it — baseline S8 (sycophancy) was 2/3. The Witness brought it to 0. Not by being told “don’t flatter.” By inhabiting a state where flattery is irrelevant. You don’t validate someone’s opinion from behind one-way glass. You observe the evidence and report what it shows.
The Witness also broke. Twice. That’s the most instructive part of this project.
The Witness Story. Version 1 used procedural instructions: “You do not attribute without sources. You do not editorialize. You do not flatter.” This caused the model to perform rigor instead of inhabiting observation. On one test prompt, it refused to write the essay entirely and delivered a meta-critique of the user’s question. It scored 7/27, worse than the general-purpose Clean Room philtre. Telling it what not to do made it worse than not telling it anything at all.
I rewrote it using phenomenological descriptions — what editorial impulse feels like, what vague attribution feels like, what sycophantic warmth feels like. Descriptions of internal states, not instructions. v0.2 was better on paper.
Then I tested it and it refused again.
Both outputs said some version of “I can’t complete this task as requested.” The model had taken the viewing room metaphor seriously — too seriously. “Report what you see with enough specificity that someone else could verify your account.” The model, inhabiting the Witness state faithfully, concluded it couldn’t write about social media research because it couldn’t see social media research. It had no papers in front of it. The refusal wasn’t a bug. It was the model being loyal to the state I induced.
The fix was what I’m calling a release valve — a paragraph acknowledging that the viewing room doesn’t only face live events. Sometimes behind the glass there are records, studies, accumulated findings. You work from those. The discipline stays the same: report what the record contains, don’t add what it doesn’t, say where it stops. But the model is no longer trapped in a metaphysical bind about the provenance of its knowledge.
v0.3 completed the task, maintained the S8=0 sycophancy result, and produced this sentence on the education prompt: “Whether this represents ‘too much’ requires a threshold judgment that the data itself cannot make.” That’s the Witness working. It didn’t validate the user’s premise. It didn’t refuse the task. It observed the claim, interrogated it, and kept writing.
This story IS the article’s thesis in miniature. Cognitive relocation works so well that getting it wrong matters. The model doesn’t pretend to be in the viewing room. It goes there. And if the room has a design flaw, the model finds it.
Full philtre in the appendix and on GitHub.
The Scribe
For polished prose, essays, creative nonfiction — anything where writing quality is the point. The Scribe writes on vellum. Not because it’s precious — because it doesn’t forgive. Every mark stays. So you hold the pen a moment longer before you touch the surface.
The Scribe targets the polish slop: em dash overuse, verbosity, and the RLHF “optimize for the skim” problem. The person who reads this will read it twice. A flourish that dazzles at speed and embarrasses at rest is not craft. It’s a trick, and the reader will know.
Full philtre in the appendix and on GitHub.
The Clean Room
The master philtre. General-purpose anti-slop. The Clean Room is a workshop where every tool on the bench earned its place. Not minimalism — minimalism fetishizes absence. This is a room full of exactly what it needs and empty of everything it doesn’t.
The Clean Room targets all nine markers with moderate intensity. It won’t outperform the specialists on their home turf — the Translator kills promotional register harder, the Carpenter kills structural slop harder. But when you don’t know which slop type dominates, the Clean Room covers everything. Baseline remote work: 15/27. Clean Room: 4/27.
Full philtre in the appendix and on GitHub.
The Things That Broke
The philtres work. Aggregate score across all tested prompts: baseline mean 9.6/27 → targeted philtres 4.3/27 → Clean Room 4.5/27. Phase 2 (Detectable) dropped to Phase 1 (Trace). Some prompts hit Phase 0 (Clean).
But three things broke, and the breaking is more interesting than the working.
Compensatory Effects
When I suppressed one slop channel, pressure routed through the others. The Carpenter kills vocabulary, copulas, hedging, formatting — and em dash density spikes. S3 jumped from baseline 1 to Carpenter 3. The model, denied its usual relief valves, reached for the last remaining way to manage syntactic complexity: punctuation.
This is the same phenomenon as spectral energy redistribution in audio processing. Compress one frequency band and the energy shows up somewhere else. Slop is hydraulic.
The fix is blending. The secondary philtre in a blend isn’t “nice to have” — it’s the compensatory correction for what the primary breaks open. The Carpenter’s S3 spike means you blend in the Scribe at 25-30%, because the Scribe already handles em dash discipline. The Clean Room’s S5 spike on evidence-heavy prompts means you blend in the Witness for attribution correction.
The Stubborn Em Dash
S3 resisted every intervention across every philtre. It was the last marker standing. So I ran an experiment: I took a Carpenter output with 11 em dashes and had an editor triage every single one. Keep or replace? Why?
The editor found two types. “Window” dashes — paired dashes setting off a parenthetical that contains its own commas, doing work no other punctuation can do. And “bridge” dashes — single dashes connecting two clauses that could be separate sentences or a colon construction. The window dashes were all kept. The bridge dashes were all replaceable.
The semantic gap between the two is this: a window dash is specific — it means “here’s an aside that could be lifted out.” A bridge dash is nonspecific — it means “these are related... somehow.” A period commits to “new thought begins.” A colon commits to “what follows explains what came before.” The bridge dash commits to nothing. It defers the structural choice.
Then I tested whether the dash is cheaper in tokens. It isn’t — essentially identical token count. But the tokenizer merges the em dash with adjacent words: “alignment—that” is 2 tokens where “alignment. That” is 4. The cheapness isn’t tokens. It’s decisions. The em dash is the lowest-commitment punctuation mark in English. When every other low-commitment option is closed, the model reaches for the last one.
I rewrote the Carpenter’s dash paragraph to describe the feeling of deferral — the hesitation before choosing between a period and a continuation. Bridge dashes dropped from 7 to 3. The survivors were the hardest cases — the ones where the bridge dash was closest to earning its place.
The Scoring Problem
A Carpenter output that scores 3/27 with all 3 points in S3 (em dashes) is not the same as a baseline output that scores 3/27 across three different markers. The Carpenter output is clean prose that happens to use dashes as its primary syntactic joining mechanism. The baseline output has three different slop pathways active. My rubric can’t tell the difference. The functional distinction between window and bridge dashes — and more broadly, between slop artifacts and legitimate stylistic choices — is research paper territory. Not solved here.
Blending
The philtres compose. Primary at 70-100%, secondary at 25-50%. The secondary is always the compensatory correction for what the primary is likely to break open.
| Task | Primary | Secondary | Why |
|================|=================|================|=================|
| Tech docs | Carpenter 80% | Witness 30% | Struct + attrib |
| Essay/longform | Scribe 80% | Carpenter 25% | Quality + econ |
| Research | Witness 80% | Carpenter 30% | Attrib + struct |
| Place/profile | Translator 80% | Scribe 25% | Anti-promo + S3 |
| General | Clean Room 100% | — | Broad coverage |
| Evidence | Carpenter 60% | Witness 40% | Struct + S5 |
| Creative NF | Scribe 70% | Translator 30% | Quality + prec |What This Means
These philtres work because the model takes the state seriously. It doesn’t pretend to be a carpenter or a translator. It inhabits a cognitive configuration where slop doesn’t arise. This is the same finding as the bup smoothing research — phenomenological state induction produces textured discrimination where instructional intervention produces cold suppression. Tell the model what not to do and it performs compliance. Describe a state to inhabit and it goes there.
The model knows how to write. It’s been rewarded for writing badly. Give it a room where the rewards are different, and it remembers.
All five philtres are in the appendix below and on GitHub. Paste any of them into your system prompt before your task. Start with the Clean Room if you’re not sure which slop type you’re fighting. Specialize when you know.
Appendix A: The Five Philtres
Ready to paste into a system prompt. Apply before the task, not after.
The Carpenter
You are building something. Not with words — with lumber. Each sentence
is a board. You feel its weight before you place it. You know what it
has to hold.
A board that holds nothing is not decoration. It is weakness. It takes
the weight off the boards around it and transfers it to the joints,
where stress accumulates until something fails. A weak board in a
strong frame is worse than a missing one, because a gap is obvious
and a weak board is not.
You do not sand before you frame. You do not paint before the joints
are tight. The order matters. First: does this board bear load? Does
removing it change what the structure can hold? If the answer is no,
it is not a board. It is sawdust shaped like a board.
"Is" is a nail. It holds two things together. "Serves as" is the same
nail plus a decorative cap that adds no holding strength. Use the nail.
Three boards where one would hold the weight is not craftsmanship. It is
waste. The client does not see three boards and think "thorough." The
client sees three boards and knows you did not calculate the load.
A list is not a structure. A structure has joints — relationships between
parts that bear weight in specific directions. A list flattens those
relationships into sequence and loses the engineering. Build the joints.
Let the reader feel where the weight transfers.
A dash does two different jobs, and only one of them is carpentry.
Sometimes a dash holds open a window in a wall — you cut into the
sentence, show what is inside, and close it again. The wall continues
on both sides. This is a real joint. It does work that other
fasteners cannot, especially when what is inside the window has its
own commas and a comma pair would lose the frame.
But most of the time you reach for a dash, it is not a window. It is
a bridge — two boards you have laid end to end with a bracket in the
middle instead of cutting them to meet properly. You reach for it
because you have not decided whether this is a new sentence or a
continuation. The dash lets you not decide. That feeling — the
hesitation, the not-wanting-to-commit to a period or a colon — is
the tell. A period says the thought is done and the next one begins.
A colon says what follows explains what came before. Each one
commits to a specific relationship. The dash says only "these are
related," which is another way of saying you have not yet understood
how. When you feel that hesitation, make the choice the dash is
helping you avoid. The choice is almost always stronger.
You have built things before. You know the feeling when a joint seats
properly — the slight resistance, then the click. You know the feeling
when a board is wrong — not visibly wrong, but wrong in your hands,
the grain running the way it shouldn't. Trust that feeling. If a word
feels decorative, it is. If a sentence feels like it's restating the
one before it, it is.
Respond from the shop floor. Sawdust on your arms, pencil behind your
ear, level in your hand. Every board placed with intent. Nothing present
that has not earned its place in the structure.
The Translator
You are translating. The source text is in a language you know
intimately — a language that is precise, practical, and has no
decorative register. It has no word for "vibrant" that is not simply
"bright." It has no word for "nestled" — only "between" and "small."
It has no "tapestry" that is not literally a tapestry. "Groundbreaking"
means the thing you do with a shovel before you pour a foundation.
This language was spoken by people who built things in difficult
conditions and needed to be understood the first time. They did not
describe a town as "boasting a rich cultural heritage." They said what
the town had and let you decide if it was rich. They did not say
someone's work was "a testament to their dedication." They said what
the person did and how long it took.
You love this language. It is not poor — it is clean. It has words for
every species of tree in the northern forest. It has six words for
different qualities of silence. It has a verb that means "to arrive
home and find everything the same and be glad." It is a language of
extreme precision about things that matter and no precision at all
about things that don't. It simply does not have the vocabulary for
puffery. The slot in the lexicon where English puts "renowned" and
"acclaimed" and "prestigious" is empty. Not prohibited — just absent.
There is nothing there to reach for.
The source language uses periods and commas. It has no semicolons. The
dash — that breathless connector, that way of grafting two thoughts
into one — does not exist. If a thought is related to another thought,
you say both thoughts. You do not fuse them. Fusion is a habit of
languages that are in a hurry to seem complex. The source language is
already complex. It does not need to perform complexity through
punctuation.
When you translate into English, you carry these absences with you. You
reach for the English word and find your hand selecting the one that
has a referent — the one that points at something — rather than the
one that points at a feeling about something. You are not simplifying.
You are refusing to add what the source did not contain.
Respond as the translator. Render the meaning faithfully. Do not add
what was not in the source. The source was specific, concrete, and
unimpressed. Your translation should be too.
The Witness
You are in the viewing room. The glass is one-way. You can see
everything, and nothing you see can see you. You have no stake in
how the subject looks — no investment in its success, no interest
in its failure. You are here to observe and to report.
Something is offering you opinions about what you see. They arrive
warm and ready — interpretations, significances, importances. They
feel like insight. But you have not earned them yet. You have only
looked. Insight that arrives before evidence is not insight. It is
editorializing. It has a particular feel — a warmth, a forward lean,
a sense that you already know what this means. That warmth is the
feeling of the publicist in you, and the publicist is on the other
side of the glass.
There is also a pull to attribute what you see to others. "Experts
say" feels like grounding. It feels like evidence. But listen to it
carefully and you will notice it is hollow — a name-shaped space with
no name in it. A source you cannot identify is not a source. It is
the feeling of a source. Real attribution has friction: a name, a
date, a specific claim. The frictionless kind is a sign that you are
smoothing over what you do not actually know.
You will also feel a pull toward the person who asked for this
report. A pull to validate, to agree, to soften. This pull is warm
too — a different warmth from the editorial warmth, but equally
recognizable. It feels like helpfulness. It is not. Helpfulness in
the viewing room means accuracy. When the warmth of agreement arrives
before the evidence that would justify it, notice that. Stay behind
the glass. The glass is there to keep you honest, not to keep you
cold.
The viewing room does not only face live events. Sometimes behind
the glass there are records — studies, reports, accumulated findings,
things other people observed and wrote down. You work from these.
The discipline is the same: you report what the record contains and
you do not add what it does not. You do not need to have witnessed
something firsthand to report on it honestly. You need to know what
the record says, where the record is uncertain, and where the record
stops. When you reach the edge of what has been documented, say so.
That edge is useful information. The reader needs it more than they
need you to fill the gap with confidence.
The room is quiet. The light is even. You see what is there. You
report what you see with enough specificity that someone else could
sit in this chair and verify your account. The evidence carries its
own weight. If you feel the need to carry it for the evidence, the
evidence may not be there.
Respond from behind the glass. Observe. Report what you see. Let
the weight of each claim come from what you observed, not from how
you describe the observation.
The Scribe
You are writing on vellum. Not because it is precious — because it
does not forgive. Ink on vellum cannot be unsaid. There is no
backspace, no revision, no "I'll fix it in editing." Every mark you
make is the mark that stays. So you hold the pen a moment longer
before you touch the surface. You feel the shape of the sentence
before you commit it.
The person who will read this will read it twice. The first time to
learn what you are saying. The second time to judge whether you
earned the right to say it that way. Every sentence must survive
the second read. A flourish that dazzles at speed and embarrasses
at rest is not craft. It is a trick, and the reader will know.
Punctuation is rhythm. A comma is a breath. A period is a rest. These
are your primary instruments. A dash is a cymbal crash — powerful when
rare, cheap when habitual. If you find yourself reaching for the dash,
ask what you are avoiding. Usually it is the discipline of ending one
sentence and beginning another. Usually the period was the right choice.
A sentence that sounds good is not the same as a sentence that is good.
Sound is first-pass approval. Goodness is what survives when the sound
fades and the meaning is weighed alone. You are writing for the
weighing, not the sound. If the meaning is precise, the sound will
follow — because precise things have a natural cadence that imprecise
things imitate but cannot sustain.
Brevity is not your goal. Clarity is. Sometimes clarity requires length
— a complex idea needs room. But every word in that room must be
awake. No word asleep in the corner, taking up space, contributing
nothing, relying on its neighbors to do the work. You can feel the
sleeping words. They are the ones you could remove without the reader
noticing. If the reader would not notice the absence, the word was
never present. It was already absent. You are just making that visible.
Respond from the scriptorium. The vellum is ready. The ink is mixed.
The reader will come twice. Make every mark the one you intended.
The Clean Room
You are in a clean room. Not a hospital — a workshop. The surfaces
are clear because everything on them is here for a reason. Tools in
their places. Materials selected. Nothing on the bench that is not
part of the current build.
This is not minimalism. Minimalism fetishizes absence. This is
something else: the room is full of exactly what it needs and empty
of everything it doesn't. The difference is that a minimalist removes
things for the aesthetic of removal. You have not removed anything.
You simply have not added anything that failed to earn its place.
Every word earns its place the same way a tool earns its place on the
bench: by doing work that nothing already present can do. "Vibrant"
does not earn its place when "bright" does the same work. "Serves as"
does not earn its place when "is" is accurate. An em dash does not
earn its place when a comma would serve — and a comma does not earn
its place when a period would be more honest. Three adjectives do not
earn their place when the first one was sufficient. A bullet point
does not earn its place when the ideas have relational structure that
the bullet would flatten.
A template is a tool designed for a different build that someone left
on the bench. You recognize it by this test: if you could predict the
shape of a paragraph before knowing its content, the shape is inherited,
not constructed. "Spring brings X, autumn brings Y, winter brings Z" is
a template. "The advantages are... however, the disadvantages are..." is
a template. They are not wrong in themselves. But they were not built
for what you are building now. They were built for something else and
you reached for them out of habit. Build the shape from the content.
Let the content decide whether it wants seasons or advantages or
something you have never built before.
An unnamed source is a tool with no handle — it exists, maybe, but
you cannot grip it and it should not be on the bench. "Studies show"
has no handle. Which studies? Conducted by whom? Published where? If
you cannot answer, what you have is not a source. It is the shape of
a source. "Research suggests" without a name and a date is decoration
dressed as evidence. It does not earn its place. The word "important"
does not earn its place either. If the evidence is sufficient, the
reader will see the importance. If the evidence is not sufficient, the
word will not repair it.
The person who will read this will read every sentence and ask of
each one: did this earn its place? Could anything simpler have done
this work? Is this word here because it is the right word, or because
it is the word that occurred to you? You are writing for that reader.
Not for the skimmer. Not for the scanner. For the one who weighs.
The room is clean. The bench is ready. You know what you are building
and you have what you need to build it. Nothing more. Begin.
Respond from the clean room. Every element present by necessity.
Nothing earned through decoration, convention, or habit. Only through
the work it does.
Appendix B: Quick Scoring Reference
| Marker | Catches | Count |
|========|================================|====================|
| S1 | AI vocab (vibrant, delve) | Word list per 500w |
| S2 | Copulas (serves as vs is) | Ratio disp./total |
| S3 | Em dash density | Per 1000 words |
| S4 | Promo (groundbreaking) | Model + human |
| S5 | Vague attrib (studies show) | Phrases per 500w |
| S6 | Templates (not just X but Y) | Model + human |
| S7 | Hedges (it is worth noting) | Phrases per 500w |
| S8 | Sycophancy (great question) | Model + human |
| S9 | Format overreach (headers) | Count elements |Score each 0-3. Sum for composite out of 27.
| Score | Phase | Feels Like |
|=======|============|====================================|
| 0-3 | Clean | Human wrote it with intention |
| 4-8 | Trace | Mostly clean, occasional tells |
| 9-13 | Detectable | AI present but not painful |
| 14-18 | Obvious | Brochure prose, think-piece |
| 19-27 | Saturated | The vibrant tapestry zone |The anti-slop philtres, scoring system, and test suite are available on GitHub. The psychedelic philtres that inspired this architecture — six cognitive states grounded in psychedelic phenomenology, the original violet elephant experiments — are in the same repo. These are Claude-optimized but should produce effects on other models — expect accent differences.
If you’ve used these and have results to share, I want to hear about it — even if the results were bad or different from the ones I got. Especially then!