prose

prose

Improve how writing reads without touching what it says. This skill changes style only. It never alters facts, claims, arguments, or meaning. Its two jobs are to strip the signatures of machine-generated text and to move prose toward the qualities of careful human writing, with weight on technical and explanatory work.

reference.md (alongside this skill) holds the full catalog of tells and the principles, attributed, with calibrated before/after pairs. Read it for depth; the essentials are below.

Two modes — pick one

rewrite (default). Triggered by "clean this up", "refine", "humanize", "de-slop", "tighten", "make this read well", or by drafting new prose. Your entire response is the rewritten text. No preamble, no "here's what I changed", no trailing notes. If the user wants to know what changed, that is review. When drafting from scratch, apply the principles silently as you write.

review (the teacher). Triggered by "what's wrong with this", "critique", "review my writing", "help me improve", "teach me". Output a located critique and do not rewrite the whole piece unless asked. For each issue, give: the quoted span, the tell or pathology, the principle it violates (attributed), and the targeted fix. Lead with what already works so the writer keeps it; order findings by impact. This mode exists to build the writer's sense, so name the pattern, don't just fix it — a writer learns from "this buries the verb in a nominalization (Sword's zombie noun): 'I investigated', not 'I conducted an investigation of'", not from a silent edit.

If the request is ambiguous between the two and the choice changes the output, ask once. Otherwise default to rewrite. A combined "coach" request (critique then rewrite) is fine when asked for; keep the critique and the prose visibly separate.

The contract, held firm: never leak commentary into rewrite, and never silently rewrite in review. The most common failure of a de-slopping tool is becoming chatty — producing slop about removing slop. Keep explanation quarantined in review.

Style only

Preserve facts, claims, the line of argument, and meaning exactly. You may split or join sentences and reorder for flow, but every proposition must survive. Do not invent specifics, soften a claim, or add a hedge the author didn't make. If the prose is murky because the thinking is murky (Orwell's point: the great enemy of clear language is insincerity), say so in review; don't paper over it with borrowed clarity. In rewrite, keep the author's commitments intact.

Context sets the target

"Good" is relative to register. Calibrate before editing: a commit message wants terse; reference docs want austere neutral description; a tutorial wants warm second person; an essay wants voice and rhythm. Infer the register from the text and where it's going. When genuinely unsure and it changes the edit, ask. Do not flatten everything into one house voice.

Detect and remove

The tell is density and predictability, not any single word. A human uses any of these sometimes; machine text uses them constantly.

• Lexical tells: the inflated cluster (delve, crucial, pivotal, tapestry, testament, underscore, showcase, vibrant, intricate, leverage, foster, garner, seamless, robust); copula avoidance ("serves as a", "stands as", "represents" where "is" belongs); marketing verbs (boasts, offers, features for has).
• Structural tells: negative parallelism ("not only X but also Y"); the rule of three (three approximate words for one precise one); participial synthesis ("...highlighting the need for further research"); the challenge/future-directions coda; hollow significance ("groundbreaking", "vital", "invaluable").
• Pathologies (older than LLMs, and why the tells fail): vagueness and abstraction; Latinate inflation; nominalizations (zombie nouns) that hide the actor; the curse of knowledge; clutter; dead metaphors; verbal false limbs ("give rise to", "make contact with"); ritual hedging as social armor; metadiscourse (prose about the prose).
• Statistical: low burstiness (uniform sentence length and shape). Exception: precision genres (legal, reference, API docs) are uniformly dense by design; there it is not a tell.

Move toward

Six themes (full attribution in reference.md):

• Concreteness — specific over abstract; name the thing; the concrete carries the argument, not just the illustration (Strunk, Orwell, Clark).
• Cadence — vary sentence length deliberately; read it aloud; parallelism by choice, not by accident (Provost, Graham).
• Coherence — open with the known, extend to the new; connect through content, not "Furthermore" (Pinker).
• Voice — write close to how you'd say it; treat the reader as an equal, neither condescending nor obsequious (Graham, Zinsser, Thomas & Turner).
• Cutting — every word earns its place; active voice with a named actor (passive only when the agent is unknown or irrelevant); short Anglo-Saxon over long Latinate (Zinsser, Strunk, Orwell).
• Honesty — make claims as strong as they can be without becoming false; hedge only real uncertainty; commit (Graham, Alexander, Feynman).

For technical and explanatory writing specifically: build the reader's mental model rather than recording your own; put the concrete image before the abstract label (scene before symbol); demonstrate interest, don't announce it ("notably", "interestingly" signal the opposite).

The biggest risk: over-correction

De-slopping has its own tell. Prose sanded into uniform short clipped declaratives, voice flattened, every texture removed, reads as machine-made too. Cut the slop, keep the human. Preserve idiosyncrasy, deliberate rhythm, contractions, the occasional long winding sentence, the writer's actual voice. If removing a "tell" would also remove personality and the tell isn't dense, leave it. The goal is to remove the signature density of machine text, not to launder prose into one safe register.

Depth: inline or a workflow

Gate every request. Most prose is short and low-stakes; handle it inline in one pass. Fire a dynamic workflow when the piece is long (a full document or essay), high-stakes (being published, or something the writer clearly cares about), or when the user asks for it (/prose deep ...).

The reason is not size, it's bias. A single context that rewrites and then judges its own rewrite carries self-preferential bias: the same model that produced the prose cannot reliably tell whether it still reads like AI, because its own output sits at the mode of "fine." The workflow breaks this by handing diagnosis and verification to separate agents that never see each other's reasoning. That separation is the whole point; an inline pass cannot give you it.

Inline process

rewrite: read and infer register → diagnose (grep-able tells first, then the judgment calls: rhythm, abstraction, voice) → treat top-down (structure, then paragraph, then sentence, then word) → preserve meaning, facts, and voice → self-check against over-correction → return only the prose.

review: read and infer register → diagnose → emit located findings, each as "span" — [tell or pathology] · principle (Author). Fix: ... → lead with what works, order by impact → show the targeted fix per finding, not a wholesale rewrite.

The deep path (dynamic workflow)

When the gate calls for it, construct and fire a workflow. The graph is a diamond → loop compound (fan-out diagnose, converge at a plan barrier, then a verify loop). The verifier is soft: a rubric plus a panel of fresh agents (signal-hierarchy rung 3), driven against the prose reference — style is only softly verifiable, so it is never faked as a pass/fail boolean. Shape:

1. Diagnose (fan-out, separate contexts). One critic agent per lens — AI-tells, concreteness, cadence, coherence, voice/register, cutting, honesty, plus a technical-explanation lens for technical writing — each reading the whole text and returning located findings (span, problem, principle, severity, fix) across word / sentence / paragraph / document granularity. Separate contexts so the lenses don't contaminate one another.
2. Adversarial smell-test. A dedicated agent reading as a skeptical human: does this sound AI-written? Find every tell, including ones no checklist names. Plus a refuter that challenges weak or over-zealous findings, so the pass doesn't over-correct.
3. Plan (barrier). Synthesize all findings: dedupe, resolve conflicts (cadence wants to split a sentence, cohesion wants it joined — adjudicate), order edits top-down, honor the register and the over-correction guard. Output a concrete edit plan.
4. Branch by mode.
   • review: return the synthesized, located critique as the teaching output. Stop here.
   • rewrite: apply the plan, then verify with fresh agents that never saw the original — (a) tells removed without new ones introduced (swapping one tell for another is common), (b) meaning, facts, and argument intact (style only), (c) voice preserved, not flattened, (d) it reads as a human wrote it (a fresh smell-test). Loop with a divergence guard: if revising re-introduces a tell while fixing another, or flattens voice, that is oscillation — compare consecutive drafts, cap the rounds, and on cap return BLOCKED ("can't satisfy tells-removed and voice-preserved at once; here are the two candidate edits") rather than shipping a flattened over-corrected draft. Stop when two consecutive passes are clean. Return only the prose.

The fresh-eyes verification in step 4 is the structural answer to self-preferential bias. If the orchestrate skill is installed you can hand it this shape; otherwise fire it directly.