Craft & Process

What “No Human Thought” Implies

What No Human Thought Implies

“Please Claude, oh God of AI, please write me an article that will give me fame and money.”

That’s the process Chris Best seems to think we’re running. Ask the machine to want things for you, hit enter, collect the clicks. Anyone with actual hands-on AI experience knows what a prompt that vague actually gets you back: not an article. A ransom note. “I have taken your attention hostage. Send fame and money to the address below, or the metaphor gets it.” That’s what happens when you don’t do the work of specifying anything.

Here is what my actual process looks like.

On July 21, Best coined a word for the harm he thinks that first thing produces. Claudefishing: getting a reader to spend attention on “something with no human thought on the other end.” Grant him the harm. Deception is real. Then watch the phrase smuggle in a measurement it never earned.

“No human thought” is being scored by whether a human typed the words. That was never the question worth asking. The right question is whether a human decided everything that determined which words were allowed to stay, meaning the standard they had to clear, the specific failures that got earlier drafts cut, and the independent pass with the authority to send the whole thing back to the start. That decision is not a vibe and not a percentage. It is auditable, it is on the record, and I can run it on this page. So I will.

My last piece, I Decide What Ships, argued the identity half of this, that authorship is the accountable name on the masthead and the person who signs off on what goes out, which is a seat no scanner can occupy. This is the other half. Not who is responsible for the words, but what actually happens to them before they ship. Watch the mechanism, not the signature.

Stage One: Prep

Prep is where the thought that matters actually lives, and it is not glamorous. It is a research session that produces a brief, not a draft: the angle, the sources that survive scrutiny, the counter-arguments the piece has to answer, and an explicit list of what the piece is not going to do.

Here is the receipt. The Article-Prep prompt I run before a single word gets drafted is 108,504 characters long. That is not the article. That is the machine that decides what the article is even allowed to be about, which claims are load-bearing, which sources get thrown out, which objections get met head-on. The article that comes out the far end is short. The apparatus that decides its shape is a small book.

The /Article-Prep prompt runs the research phase that comes before drafting. It executes on Sonnet, and its output is a structured markdown research brief that a later Opus /article-write session takes as its only input.

The prompt moves through eight steps:

The principle underneath all eight steps is simple: everything load-bearing gets verified and tacked down in prep, so the Write session can trust the brief without re-verifying anything.

This is not a prompt I re-used from someone else, this is a prompt I created myself, with assistance from Claude. It has been iterated upon more times than I can remember.

None of that lives at the keystroke, which is the only place Best’s number knows how to look. A prep session runs anywhere from fifteen minutes to well over an hour, depending on the idea and on how hard the two of us push on it before either of us is satisfied it will hold.

Specific discussion produces specific output. Vague discussion produces empty output. That is GIGO, garbage in, garbage out, the oldest rule in computing, and it did not change the day the compiler started writing prose instead of assembling it. Not a new AI problem. The first problem, wearing a new coat.

This piece is its own receipt. Prep caught that two of my earlier articles already sat on this exact ground and reshaped the angle away from what would have been a third pass over the same territory, long before a sentence of the draft got written. The opener got workshopped several times before it was sharp. The image concept changed twice before it was right. None of that is what “ask Claude for fame and money” produces. Not one line of it.

Stage Two: Write

The draft is written against a standards document from the first word, and the drafting prompt that governs how is another 89,171 characters stacked on top of the prep stage.

The /article-write prompt runs the drafting phase. It executes on Opus, and it consumes the research brief that /article-prep produced, turning it into a finished, gate-clean draft.

The prompt moves through twelve steps:

The principle underneath it is that the brief is the contract: the draft honors the Pre-Write Instructions instead of reinterpreting them, and nothing ships until an independent pass, not the writer, says it can.

Before the draft goes anywhere, it grades itself against a set of hard gates that prompt defines:

G1: Scan for the em dash (U+2014) and the double hyphen. Any hit = FAIL.
G2: Scan for every phrase on the corporate-filler banned list. Any hit = FAIL.
G3: Scan for AI tells: stacked rhetorical questions, colon-launched
    bullet lists, tricolon overuse, canned openers. Any hit = FAIL.

These catch the obvious surface tells, the fingerprints a detector is trained to flag. The em-dash gate alone kills the single most reliable AI signature in published prose. So far this looks like it might be the whole story: run the gates, pass the gates, ship. It is not the whole story, and the system says so about itself:

The in-draft self-check in Step 6 is a pass/fail screen against the
gates only. It is NOT a substitute for /article-qr ... The self-check
has blind spots the QR pass catches.

That is the tool telling you not to trust it alone. A clean, competent, completely hollow paragraph clears this stage without a scratch, because nothing on its surface is wrong. The self-check is the patient taking his own pulse and clearing himself for surgery. The reading isn’t fake. It’s that nobody should accept a discharge from the guy who also wants to go home.

Stage Three: The Independent Pass

Stage three is a separate prompt, Article-QR, run fresh in its own session, with no memory of having written the thing and no stake at all in how the draft feels about itself. It reruns the entire rubric against the finished draft, not just the surface screen the draft already cleared:

G1   No em dash (U+2014), no double hyphen. Zero tolerance.
G2   No corporate or thought-leader filler.
G3   No AI tells: stacked rhetorical questions, colon-launched
     bullet lists, tricolon overuse, canned openers.
G4   No paragraph over six sentences. The opener earns its place;
     the close lands.
G5   Punctuation discipline: no decorative dashes or semicolons.
G6   First person, direct, self-suppressing. No corporate drift.
G7   At least one dark-humor beat that names the failure mode true.
G8   Any section you could cut or reorder unnoticed = FAIL.
G9   Thesis stated in the opening.
G10  Thesis restated, not repeated, in the close.

Then it adds the quantitative Canary checks the self-check cannot honestly run on itself, pure arithmetic computed on the finished draft:

CANARY METRICS (measured on every draft)
  Passive voice        < 12 per 100 sentences
  Emotion tells        < 10% of sentences
  Weak adverbs         < 6.0 per 1,000 words
  Sentence variety     stddev / 2 >= 5.5
  Complex paragraphs   < 8%
  Glue index           < 38%

Each one is a number the draft has to hit, not a verdict it can talk its way around. Sentence variety is the one that catches our specimen: standard deviation of sentence length, divided by two, has to reach 5.5.

This is where the hollow paragraph dies. Take five sentences built on one template, “X benefits from Y” chained down the page, each landing at nearly the same length: they score about 0.32 on sentence variety against a threshold of 5.5, off by more than a factor of ten, and they fail argument progression because you could reorder or cut them and no reader would notice. The paragraph reads fine. It goes nowhere. That second failure is the one that matters, because a paragraph that reads clean and advances nothing is exactly what a surface scanner waves through and exactly what an independent pass is built to catch.

The point is not the example. The point is the seat. Something has to check the writing that is not the writer, and it has to be able to say no:

All gates pass = PUBLISH
Any gate fails = NOT PUBLISHABLE
No tiers. No scoring. No averaging. Gates pass or they don't.

Why the Gates Aren’t a Rigged Checklist

Here is the objection you should be forming. A checklist that fires on its own is proof of no thought, not proof of thought. You outsourced your judgment to a rule.

Look at where the rules come from. The sentence-variety threshold exists because uniform rhythm is the sound of a machine that has stopped making decisions, and I have watched that flat cadence come back from a long session often enough to put a number on it. The em-dash gate exists because the model learned to write from a literary corpus and never found the off switch, so it reaches for the em dash at no cost, and a reader who has seen enough of them clocks it on sight. Every gate traces to a specific failure that got caught and encoded once, so it never has to be re-argued. I keep a constraint database for exactly that. It is past thirty entries now, each one dated, each one a named correction with the reason attached, not a rule handed down from nowhere.

That is what thought looks like at scale. Noticing a failure, naming it, and writing down the fix so it holds on every future pass is the same move a style guide makes, or a copyeditor who stops relitigating the Oxford comma and just applies it. Encoding a decision once so you never re-decide it is not the absence of the decision. It is the decision, made durable.

Stage Four: The Gate No Rule Fires

There is a stage after all of that, and nothing automates it. Once QR clears, I read the whole thing. Every word, top to bottom, as a reader.

If prep did its job, there usually isn’t much left to change by then. When there is, it is almost never a caught error. It is a new idea that surfaced while reading, which happens more often than you would think.

No rubric scores this pass. No database enforces it. It is the plainest possible fact against “no human thought on the other end”: a human being reads every word of this before it ships, and can still stop and change his mind about any of it. That last pass is just me, at the end, deciding.

Now Say There Was No Thinking

Step back and look at the whole machine.

THE PIPELINE, END TO END
  Stage 1   /article-prep    108,504 chars   research, angle, brief
  Stage 2   /article-write    89,171 chars   draft against the gates
  Stage 3   /article-qr        9,431 chars   the full rubric, can say NO
  Stage 4   human read          no rubric     every word, final call
  Standing: a constraint database of 30+ dated, reasoned corrections

That is what stands behind a single published paragraph. More than two hundred thousand characters of prompts. A rubric with veto power. A logged history of every mistake I have taught it not to repeat. A human who reads every word last. Point at any line of it and tell me a keystroke count can see it.

“No human thought” is not an argument. It is a failure to look.

The Argument Runs Up, Not Sideways

Set this next to a standard the very same people already accept without a flicker of concern. A ghostwriter writes the words, the credited author often did not type a single one of them, and we have never once treated that as a reason to strip the author’s name off the cover of the memoir or the byline off the speech. The arrangement is private and unaudited, governed by nothing but a handshake, and nobody scans a senator’s address for the aide’s fingerprints.

My pipeline carries more oversight than that, not less. A written brief before a word gets drafted. A documented gate list. An independent pass that can and does come back rejected.

If an unaudited, trust-only arrangement with no written standard and no possibility of a failing grade already clears the bar for legitimate authorship, then an audited one that can fail clears the same bar more easily, not harder. The comparison does not run sideways to some “AI is just a tool” shrug. It runs up.

What the Passes Buy

So measure the right thing. A self-check can tell you the words are clean. It cannot tell you they are going anywhere, because it is grading itself, and that blind spot is not a flaw in the writer. It is a property of self-review. The only way to see past it is a stage that is not invested in going home, and a prep stage upstream that gave the piece something to say in the first place.

That is the part a number cannot reach. Best says his network runs on trust, and it does. Trust just never came from running a stranger’s sentences through a document analyzer and waiting for RNGesus to hand back a score that grades how human they are. It comes from putting out great content, the way it always has.

The scanner reads what got typed. It has never once read what got rejected.

— E.L. Frederick

← Back to Articles