22 Claude-prose Patterns
A Catalog of Stylistic Attractors in Generated Texts
Few posts on AI age well. Those commenting on current features, certainly not. That conviction stopped me from continuing the work on a draft of this post last year. All I did was cobble a quick taxonomy of overused words:
But those were concrete, fixed words — an easy target and less interesting than the patterns of metadiscourse and cheap rhetoric. This latter category is the focus of this post.
At first, I thought they would be fixed quickly. Yet half a year has passed, and they are not. Are they so difficult to fix, or are they considered a feature rather than a bug? I thought it was certainly the former, but the companion essay (I’ll put the link below) tells me it might be the latter. And there is a third explanation. That they are simply attractors.
I prefer using Anthropic models over OpenAI's, but I have to admit that when it comes to style, Claude leads in overuse of metadiscourse and cheap rhetoric.
It could be that, as a non-native English speaker, it’s easier for me to smell some clichés simply because they don’t come naturally to me (admittedly, some sneaked in recently). The overuse of lexical tics like “genuinely” and “structurally” is genuinely and structurally annoying but not as much as the rhetorical schemes. My top list includes the contrastive binary, which takes the form of “no X but Y” or “X is not Y; it is Z,” like in "not decoration but error-prevention" (this specific one is also a self-reflexive defense in its context, as you’ll see if you read the companion essay), and the significance-signaling as in “this matters because.” When actually counting, it turned out that these were indeed the most frequent offenders.
So, over the last six months, Sonnet, Opus and now Fable have consistently fabricated an elaborate, tightly woven tapestry of stylistic mannerisms. I thought it might be worth compiling a catalog of Claude rhetoric signatures. And who should I ask for help? Well, Claude, of course.
The experiment revealed more than I hoped for.
I first asked ChatGPT to compare the discourse patterns of GPT and Claude. Here’s one sample from the response, to give you an idea:
Claude-like
heavy framing before content lands
“this matters because…”, “what’s going on here is…”
frequent meta-explanations
GPT-like (typical default)
more direct statement-first structure
explanation comes after assertion, or not at all
less “pre-justification”
Effect: Claude reduces interpretive risk; GPT increases informational speed.
Then I used the GPT output as input to Claude, with additional instructions to make a rigorous study. Claude produced a serious article and tried but failed to avoid the overuse of clichés. So then I asked it to mark the stylistic mannerisms in its own essay about AI-generated prose, based on what had been initially cataloged as part of the exercise. Then I pointed to omitted patterns and to misclassified ones. After a few such iterations, the catalog took shape, and the essay's annotation improved.
The catalog has 22 patterns (down from 30 since eight did not match my findings). It is structured like this:
[<Acronym as index>] <Pattern name>
Examples: <examples mainly from the Claude essay >
<description>
It should be best read side by side with the essay. The essay itself is not sent by email.
The most frequent patterns are the contrastive binary [CB], mirrored-clause symmetry [MCS], and the aphoristic ender [AE]. CB and AE are often combined, creating a compound pattern of an aphoristic ender built as a binary. Patterns such as significance-signaling [SS], meta-signposting [MS], and others are consistent with findings that AI-prose, not just Claude-generated prose, tends to over-explain.
You can easily test it. When you point to an error or something missing, if Claude agrees, the usual structure of the first paragraph is: <something admitting you are right><pointing out what is still valid>, and ending the paragraph with a variant of “But that's not all it's doing, and the surplus is what you're pointing at,” which is a combination of anticipate-and-rebut [ARR] and contrastive binary [CB] (see the descriptions below in the catalog section).
Why are these patterns so stable and predictable?
They are attractors. In dynamical systems, an attractor is a state things keep sliding back toward no matter where they start. LLMs work the same way. No matter how different the prompts are, the responses, while different in content, come in a metadisourse package that tends toward reusing the same stylistic formulas. A few shapes are simply more probable than the rest as ways to round off almost any “thought.” Most likely, their probability also increases through reinforcement-learning loops, in which many people find a heavily signposted and overexplained response more helpful and understandable.
The use of an LLM is part of the reader's OODA loop. You take in the words, you work out what they mean and how they fit what you already know, you decide what the writer is getting at, and you carry on. The second “O,” orient, is when you have to figure out how to take a sentence. For example, is this the main claim or an aside; is this hedge a real limit or just politeness? It seems that Claude tends to take up more of the “orient” effort than other models. Every “this matters because” and “the deepest point is” is an outsourcing of your orientation, a convenience that has its price. The result is often more understandable (sometimes the effect is the opposite — it sounds rather dense and cryptic), but it limits other, potentially more useful interpretations.
Raising that awareness is probably the catalog's main benefit.
Another potential use of the catalog is as “manual” AI detection, but that’s not its purpose. The use of AI proves no lack of quality, just as its abstention proves no presence of it.
A better use would be as a general instruction (maybe only a subset, for cost efficiency) or as a SKILL.md. This would not solve the problem of the lack of diversity. At best, it will shift the homogeneity to less annoying basins.
The Catalog
[AE] The aphoristic ender
Examples: “a stance, not its absence”; “leaves fingerprints, and fingerprints can be counted”; “the dense version is the courtesy”; “it is relocated to the writer”; “not a property of the prose”; “evidence of training, not of virtue.”
Landing a paragraph or section on a compact, quotable epigram. Generally desirable but in Claude-generated prose is formulaic and overused. Often done via CB.
[AHM] The reflexive AI-humility move
Examples: “this essay is written by one of the systems under examination …”; “Most of the above is structured impression, and could be wrong”; “is a small instance of that, and no evidence of anything finer”.
Claude flagging that it is itself a language model, or that its claims could be wrong, as a gesture of modesty.
[ARR] The anticipate-and-rebut reversal
Example: “as though it carried no stance. It carries one”. (in this case, also a CB)
Expressing an implied opinion then quickly dismissing it, often as a brief, abrupt sentence fragment.
[CB] The contrastive binary
Examples: “not a style but an attractor”; “not the output of a controlled head-to-head”; “is not the neutral midpoint – it is one option among three”; “a stance, not its absence”; “not decoration but error-prevention”; “is not abolished; it is relocated”; “not a property of the prose”; “is one such declaration, not a neutral description”; “evidence of training, not of virtue”; “of central tendency, not of kind”; “on impression rather than measurement”
“not X but Y” / “X, not Y” manufactures a clean opposition and resolves it in the same breath. The diplomatic variation using “rather than” is the same pattern.
[CCC] The clean-consequence connector
No examples in this essay but a common Claude pattern. Here are some external examples: “The cost the essay described comes straight out of this”; “falls out of”, “follows directly”
A pattern that lends an air of inevitability not earned by the argument itself.
[CDF] The candor flag
Example: “the honest answer”.
Framing what follows as especially frank.
[CF] Contribution framing
Example in the essay: “supplies the other half”. External example: “it pins down something the earlier explanations left open”.
Presenting the point as completing or filling a gap that earlier accounts left, which makes the contribution look larger than it is.
[CL] Confidence by litotes
Examples: “not difficult to specify”; “blind tagging is not optional”.
A litote states a thing by negating its opposite. Why “confidence”? The understatement implies the reverse. The statement is neither questioned nor merely asserted. It is presupposed, completely excluding the possibility of questioning it.
[CP] The corrective pivot
Example: “It would be wrong, though, to call the difference padding”.
A pattern of staging a small debate with itself.
[CR] The colon-reveal
Examples: “not a style but an attractor: the center of gravity of a distribution” (the part before the colon is already a CB); “a measurable claim: markers per thousand words”.
A setup clause, a colon, then a tidy payload. CR builds small moments of suspense, often when they are not needed. A similar pattern is the suspense hook [SH].
[DT] The deflating tail clause
Examples: “and no more”; “blind tagging is not optional”; “and no evidence of anything finer”.
A short theatrical qualifier tacked onto the end of a sentence to look modest or precise.
[MCS] Mirrored-clause symmetry
Examples: “High-metadiscourse prose runs the channel at high redundancy; thinned prose runs it at high density,” “risks quantity in the upward direction” / “risks quantity in the downward direction”; “for someone who has never priced a bond … for someone who prices them daily …”; “The first failure wastes … the second misleads …”; “cannot minimize the risk of one without raising the risk of the other”; “Visible hedging may reflect … visible confidence may reflect …”; “the standing argument for it” / “the standing argument against it”; “over-count its hedges and under-count the other’s”.
Two clauses in one grammatical frame, set side by side (often separated by a semicolon), with only a couple of slots swapped to convey contrast or a second case. This pattern is similar but wider than CB, which uses only a single “not X but Y.”
[MS] Meta-signposting about the document itself
Examples: “Four caveats belong at the front”; “below I try to specify”; “any sentence below”; “One qualification carries through the rest of this piece”; “kept separate from the functional account above”.
That’s narrating the structure rather than just having one. Often useful, but Claude is consistently overusing it.
[RF] The reframe
Examples: “Better posed:”; “the harder skill is usually”.
This is a funny way of taking control of the framing. Claude repositions the question as a setup for its preferred version of it.
[RG] The restatement gloss
Examples: “in other words,” “put differently”
The RG announces a paraphrase of something already said. Again the problem is the overuse.
[RH] Reflexive hedging
Examples: “almost”; “tends to”; “with few exceptions”; “roughly”; “largely”.
Calibration uncertainty tracking. Not bad per se, but applied uniformly, it reads as a verbal reflex.
[SDA] The spaced-dash aside
Examples: “– the prose each model produces when a prompt specifies a task but not a register”; “– it carries no fact, only an instruction about how to read what follows”; “is not the neutral midpoint – it is one option among three” and more.
Heavy reliance on dash-flanked parentheticals for asides, qualifications and lists.
[SH] The suspense hook
Examples: “has a name”; “the cleanest organizing idea is this.”
Withholding a term or claim for one beat to create some suspense. Yet another cheap drama pattern.
[SK] Stakes-raising
Example: “because they shape everything that follows”.
Inflating the importance of what is about to be said. The payoff rarely matches the build-up.
[SRC] Self-ranking one’s own claims
Examples: “most important”; “cleanest organizing idea”; “The deepest point in the usual comparison is”; “a deflationary point”.
Telling the reader which of its points is the profound one, rather than letting them judge.
[SS] Significance-signaling
Examples: “That reduction is useful, because”; “This matters for the comparison, because.”
Significance-signaling tells the reader that what follows is significant, rather than letting it be significant. (Claude initially called it “framing throat-clearing”, and GPT “meta-justification”)
[VP] Validate, then promise precision
Examples: “is correct, and it can be made precise”; “is right, and this is its exact form”.
First, a self-issued stamp of correctness and right after, in the same sentence, a promise that an improvement will follow. The “more precise” can be left to do its own work without being ratified and pre-announced first.