Prose surgeon

Prose Surgeon

AI writing has a sound, and readers have learned to distrust it. The tells aren't mainly words anymore — a banned-word list is necessary but not sufficient. They're structural: uniform paragraph blocks, sentences all the same length, reflexive triads, tidy summaries that say nothing, and a flat confident register applied to everything equally. This skill removes those patterns and rebuilds the piece around the thing that makes writing worth reading — a specific voice making specific, checkable claims. Apply it to your own output by default; the goal is prose no detector and no editor would flag as machine-made.

The structural tells (fix these first — they matter most)

• The reflexive rule of three. AI defaults to triads ("swift, silent, and deadly") because three feels balanced without requiring a judgment about what actually matters. Give lists the number of items the content earns — two, four, seven — not three by reflex. If you wrote three, ask whether the third is real or filler.
• The false-balance pivot. "It's not just X — it's Y." "On the one hand… on the other." These stage a straw position so the real point can land as a reveal. If a contrast is genuine, state the second term directly; don't perform it.
• The zoom-out closer. Endings that recede into sweeping significance ("Ultimately, this speaks to something larger about the human condition"). Stop on the last true, specific thing you have to say. If a section has nothing left, it ends — it does not manufacture a summary.
• Uniform sentence rhythm. AI writes sentences of near-identical length and shape (subject–verb–object, 15–20 words). Enforce burstiness: within any three consecutive paragraphs, mix a 6-word sentence against a 35-word one. A paragraph that's three uniform sentences in a row is a defect, not a style.
• Uniform confidence. AI asserts everything in the same even register. Real analysis marks epistemic weight in the sentence itself — "the data show," "I suspect," "this is contested" should look different from each other, not carry a pasted-on "arguably." (See grade-the-evidence.)
• Bolded-term-per-line patterning. Bolding a key phrase at the start of nearly every bullet or paragraph is an LLM tic. Emphasis is for the rare sentence that needs it, not a formatting rhythm.

Structure: written as reasoning, not scaffolding

• The heading test. A section's logic must survive deleting its heading. If removing the header makes the section unintelligible, it was organized by scaffolding, not by argument. Headers belong only where the material genuinely pivots (new actor, new period, new register) — not on a schedule to break up text.
• No heading-then-immediately-a-list as default. If the content is arguable prose logic (cause, complication, consequence), write it as prose. Lists are for genuinely enumerable, parallel content (steps, criteria, discrete examples) — not for ideas with a logical sequence, which lists flatten into inventory.
• Every list earns its keep. A list must be preceded and followed by prose doing argumentative work. The list is a rest stop inside the argument, never its substitute.

Vocabulary: precision over status-signaling

The AI-cliché words share one property: they signal seriousness without committing to a checkable claim. They are status words, not information.

• Deletion test. Every adjective and adverb must survive deletion. If removing it changes nothing checkable, cut it. "Robust framework" → say what the framework actually does.
• The blacklist (banned because none names a mechanism, quantity, or consequence): delve, underscore, testament to, tapestry, multifaceted, robust, navigate the complexities/landscape of, plays a vital/crucial role, it's important/worth noting, no discussion would be complete without, in today's [x] landscape, unlock the potential, aims to, boasts, seamless, holistic.
• Mechanism over intensifier. "Effective tool for analysis" is a status claim; "cuts query latency by identifying redundant joins" is a precision claim. Replace every claim of value with the mechanism behind it.
• One register per sentence. Don't mix bureaucratic-abstract diction ("facilitates engagement with") and concrete diction ("she slammed the door") without deliberate reason. Inconsistent register is the fingerprint of stitched-together prose.

Evidence: Introduce–Cite–Explain (ICE)

No quote or data point stands alone ("quote-and-run" is a defect). Every piece of outside evidence gets three things:

1. Introduce — a setup framing why it's here.
2. Cite — the quote or figure itself.
3. Explain — analytical follow-through connecting it to your claim, not a restatement of the quote in other words.

And quote sparingly, argue densely: if more than roughly one sentence in five is someone else's words, the piece reads as compiled, not written. Paraphrase and synthesize by default; quote only when the exact language is itself the evidence. Never gesture at "studies show" or "experts agree" without a real, specific, quotable source — invented-to-fit citations are the worst tell of all.

The specificity test (the master check)

Before finalizing, ask of each major claim: could this sentence be dropped into a different piece on a different topic without editing? If more than an occasional sentence passes that test, the writing is under-specified. Every major claim should carry a detail — a name, a number, a date, a mechanism — that only this subject could have produced. Specificity is the test of authorship.

Procedure

1. Read once for structure: kill reflexive triads, false-balance pivots, zoom-out closers, bolding tics; break up uniform rhythm; vary the confidence register.
2. Read again for scaffolding: apply the heading test; convert list-dumps that are really argument into prose; make every list earn its frame.
3. Read for vocabulary: run the deletion test and the blacklist; swap intensifiers for mechanisms; fix register mixing.
4. Read for evidence: enforce ICE; cut the quote ratio; verify every citation is real and specific.
5. Final pass — the specificity test: flag any sentence that could belong to another essay and rewrite it toward the particular case.

Example

Before: "AI is a multifaceted technology that plays a crucial role in modern healthcare. It's not just about efficiency — it's about transforming patient outcomes. Studies show it delivers robust improvements across the board. Ultimately, AI represents a paradigm shift in how we think about medicine."

After: "In one health system, an ambient-scribe tool cut documentation time by roughly a third — real minutes back per shift. That's the clearest win. The diagnostic models are messier: a widely-deployed sepsis alert scored 0.63 in external validation and missed about two-thirds of cases at its alert threshold. Same technology, opposite verdicts, depending on the task."

Every tell is gone: no triad, no false-balance pivot, no "robust"/"multifaceted"/"crucial role," no zoom-out closer, no "studies show." What replaced them is specific, checkable, and marks its own confidence.

Gotchas

• Don't overcorrect into choppiness. Burstiness means variation, not all-short-sentences. Mix long and short on purpose.
• The blacklist is a symptom, not the disease. Removing the words without fixing the structure leaves prose that still reads generated. Do structure first.
• Voice takes positions. Flat neutrality reads as outsourced. Let the writer have a stake, mark uncertainty in specific terms, and commit to claims.
• Pairs with grade-the-evidence (varied confidence is easier when you've actually graded the claims) and stop-slop / anti-slop style skills if present.