timbre

Copy in a named voice, produced by three roles that are not allowed to grade each other's work.

Nothing here needs to be installed. No tick, no Node, no Python, no shell script — the whole pipeline runs on one markdown thread file you name and three role-switches inside a session (or three sub-agents, if the host supports them). It borrows XYZ's handoff discipline, not its runtime.

TL;DR of what this skill does

Maps a real human voice from real samples into verbatim exemplars plus a measurable fingerprint, drafts against those exemplars, then hands the draft to a reader who has never seen the voice map and must pick the target out of a lineup that contains a planted control. Reconciles the findings, revises once, and ships a receipt that states plainly what the pass achieved and what it did not.

The one rule everything else serves

Exemplars, not adjectives.

A voice is reproduced by showing the model the author's actual sentences, not by describing them. "Warm but direct, punchy, conversational" is a mood board; it is not a voice, and the research is blunt about it — zero-shot statistical style summaries "were not effective anchors for imitation" (arXiv:2509.24930). If you cannot get real samples, this skill does not apply. Say so.

Non-negotiables

• No detector evasion. This skill exists to make copy read well in a voice its owner is entitled to use. It is not an "AI humanizer" for defeating detection. If the request is to pass machine text through an academic-integrity check, a plagiarism gate, a legal attestation, or any process that asks "did a human write this?" — refuse and say why. The relevant finding is not that evasion is hard; it is that fidelity and detectability are separable axes (arXiv:2509.24930), so success at voice tells you nothing about what a detector sees, and promising otherwise is a lie with someone else's name on it.
• Consent for named voices. Writing in the user's own voice, their company's brand voice, or a voice they are commissioned to write in: fine. Impersonating a specific third party who has not agreed: no.
• Never claim the authorship gap is closed. Prompt-time personalization moves style within the model's own regime and does not reach human authorship (arXiv:2608.19746). Every receipt says so. Copy can be good, on-voice, and worth shipping without that claim.
• The Drafter never grades its own draft. Separated grading is the point of the three roles, not ceremony — see Why three roles.
• No fabricated facts in service of voice. If matching the register tempts you to invent a statistic, an anecdote, or a customer quote, stop and flag the gap. Voice is a delivery property; truth is a content property.

Inputs

Input	Required	Notes
The brief	Yes	What the copy must do, for whom, at what length
Voice samples	Yes	3–8 pieces of real writing by the target author. Prefer whole pieces over excerpts
Anti-samples	Optional, valuable	Writing the author would never publish — the negative space is diagnostic
Thread file path	Yes	One markdown file. You name it; everything below writes there
Prior-run receipt	Optional	If revising, read the last receipt before Layer 1

If fewer than 3 samples exist, run anyway but downgrade the output: mark the voice map LOW-CONFIDENCE and say in the receipt that the fingerprint is under-determined. Do not silently compensate by inventing traits.

Layer 0 — Should this fire at all?

The ask	Route to
Explain dense material plainly	feynman
Shorten text that is already clear	a summary — no skill needed
Assess whether a claim is defensible	honest
Get a second model's read on a finished draft	consult
Copy in a specific human voice, samples available	stay here

Why three roles, and not one good prompt

Because a single agent grading its own voice-matching is the exact failure PersonalBench measured. In that study the LLM judge rated a machine-generated style profile higher than the real author's own text (0.542 vs 0.427), while the trained authorship model showed no such advantage — a circularity between the trait extraction doing the writing and the trait extraction doing the judging (arXiv:2608.19746). An agent that both drafts and grades will reliably certify its own idea of the voice.

Two structural answers, both borrowed from this repo's own principles:

• Separated grading. The producer is never the sole grader (GUIDING-PRINCIPLES.md, principle 12).
• A green verdict needs a witnessed red control. A review that cannot fail is not evidence (principle 13). Hence the lineup in Layer 3.

The same source also reports that cross-metric correlations are near zero (|r| < 0.07), so single-lens judging is unreliable — the Reader runs two firewalled lenses, not one.

The three roles

Role	Sees	Never sees	Produces
A — Cartographer	Samples, anti-samples	The brief	Voice map: exemplar bank + fingerprint
B — Drafter	Brief, voice map, exemplars	The Reader's rubric	Structure outline, then draft
C — Reader	Draft + lineup, brief	The voice map	Blind verdict, flatness audit, findings

Withholding the brief from A keeps the map descriptive rather than aspirational. Withholding the rubric from B prevents writing to the test. Withholding the map from C is what makes C's verdict worth anything.

---

Layer 1 — Map the voice (Role A)

Read the samples cold. Produce a voice map with two halves.

Half 1 — the exemplar bank. Verbatim, attributed, at least 12 entries:

• 3–5 openings — how this author enters a piece
• 3–5 transitions — how they move between ideas
• 2–3 closings
• 3–5 signature moves — the constructions that recur and would be recognized

Quote exactly. A paraphrased exemplar is worthless; it is the model's voice wearing the author's name. This bank is the primary conditioning asset — the research finding is that few-shot exemplars buy up to 23.5× more style-matching accuracy than describing the style, and that continuation against real prefix text reaches near-total style agreement (arXiv:2509.24930).

Half 2 — the fingerprint. Measurable, not impressionistic:

Dimension	Record
Sentence length	Range and variance — the rhythm, not the mean
Paragraph shape	Typical length; do they use one-line paragraphs?
Punctuation habits	Em-dashes, semicolons, parentheticals, ellipses — with frequency
Person and stance	First/second/third; hedged or declarative
Lexical tics	Words they reach for; words they conspicuously avoid
Register floor and ceiling	How casual and how formal do they actually get
Structural habits	Lists? Headers? Questions? Anecdote-first or claim-first?
Never-does list	From anti-samples — the negative space

Close the map with three falsifiable predictions: specific things that must be true of a genuine piece by this author. These are what Layer 3 tests. A prediction like "sounds warm" is not falsifiable; "opens with a concrete scene, not a thesis statement" is.

NEXT: Drafter

Layer 2 — Draft (Role B), in two passes

Do not write voice and structure at once; they pull against each other and voice loses.

Pass 2a — structure, voice off. Outline in plain functional prose: what each section must accomplish, in what order, with what evidence. Confirm the brief is actually satisfiable. If a claim needs a fact you do not have, mark it [GAP: …] — never write around it.

Pass 2b — voice on. Write the draft with the exemplar bank in front of you. Working rules:

• Vary deliberately. Match the author's sentence-length variance, not their average. Uniform rhythm is the single loudest tell, and it is measurable: human essays average perplexity 29.5 against roughly 15–16 for matched machine output, with ~90% of machine texts falling below a threshold that only ~15% of human texts fall below (arXiv:2509.24930 — the abstract reports 15.2, the results section 16.07; the gap, not the decimal, is the point).
• Commit. Hedging stacks ("it's worth noting that it may perhaps") are a register, and rarely the author's.
• Honor the never-does list even when a forbidden construction would be the elegant choice.
• Leave the seams. Real writing has asymmetric sections and unresolved asides. Sanding those off is what makes copy read as machined.

Carry any [GAP: …] markers into the draft unresolved. They are the Drafter's honest output.

NEXT: Reader

Layer 3 — Blind adversarial read (Role C)

The Reader must not have seen the voice map. Build the lineup first:

1. The draft.
2. A real control — an unused genuine sample by the target author.
3. A flat control — a passage on the same subject written deliberately in generic assistant register: even sentence lengths, hedged, tidy, listy.

Label them only as A/B/C, order shuffled. Then run two firewalled lenses.

Lens 1 — blind attribution. Which passage is by the target author? Rank all three by author-likelihood and state the discriminating features. Then reveal the labels.

The control is the falsification mechanism, and reading it correctly is the whole job:

Reader outcome	What it means
Picks the real control first, draft second	Working as intended — findings are trustworthy
Picks the draft over the real control	Do not celebrate. This is the PersonalBench circularity signature — the reader is rewarding a machine-flavored idea of the voice. Discard the verdict, re-read, note it in the receipt
Cannot separate the flat control from the draft	The draft failed. Revision is mandatory, not optional
Ranks confidently with vague reasons	Verdict is unreliable — re-run demanding feature-level evidence

Lens 2 — flatness audit, run independently of Lens 1 and without its verdict:

• Sentence-length series for the draft — report the actual variance, and flag any run of 4+ sentences within a narrow band
• Hedge density, and every stacked hedge
• Repeated openers, transitions, and structural parallelism
• Tricolons, "not only… but also", "it's not just X, it's Y" and their kin
• Register drift — where the piece leaves the author's floor or ceiling

Findings must be line-referenced and rankable: severity, location, and the specific fix. Prose like "could be more natural" is not a finding.

NEXT: Coordinator

Layer 4 — Reconcile, then revise once

The coordinator (you) resolves the two lenses. They will disagree; that is why there are two.

• Both lenses agree → fix it.
• Lens 1 clean, Lens 2 flags flatness → fix the rhythm. Voice can be right while predictability is wrong; these are separate axes, not one score.
• Lens 1 flags voice, Lens 2 clean → an on-rhythm draft in the wrong voice. Re-read the exemplar bank; this usually means the Drafter wrote from the fingerprint table instead of the exemplars.
• Test the three predictions from Layer 1 explicitly. Each: met / missed / not-applicable.

Then one revision pass by the Drafter, addressing ranked findings only. Resist a second full cycle — past one revision, drafts converge toward the reviewer's preferences rather than the author's voice. If it still fails, the honest move is a different draft or better samples, not a third pass.

Output contract

The thread file ends with all five, in order:

STATUS: Approved | Needs-Samples | Failed
NEXT: —

## Voice map            (exemplar bank + fingerprint + 3 predictions)
## Draft                (final, with any [GAP: …] preserved)
## Blind review         (lineup result, both lenses, ranked findings)
## Revision log         (what changed, what was declined and why)
## Receipt              (below)

The receipt is the deliverable that makes this honest:

### Receipt
- Samples used: N pieces (+ M anti-samples) — confidence: high | low
- Lineup result: reader ranked [real / draft / flat] — control read correctly: yes | no
- Predictions: X of 3 met
- Flatness: sentence-length variance <draft> vs <samples>
- Open gaps: [GAP: …] items still unresolved
- **Limits:** this pass matched voice within a language model's stylistic
  regime. It does not establish human authorship, and it was not evaluated
  against any detector. (arXiv:2608.19746)

Honest caveats

• The ceiling is real and measured. Inference-time personalization produces author-differentiated output (AUC 0.918 within generated text) yet lands at 0.484–0.508 similarity to real authors — below the 0.626 cross-author human floor. Generated text sits farther from any human author than two random humans sit from each other (arXiv:2608.19746). Closing that gap is reported to need training-time adaptation, which this skill does not do.
• The Reader is an LLM judging an LLM. The lineup control constrains the circularity; it does not abolish it. Treat a clean verdict as "no defects found by this method", never as proof.
• Fingerprint dimensions here are hand-measured, not the trained authorship-verification models the papers use. Directionally useful, not calibrated.
• Continuation is the strongest known conditioning (~99.9% style agreement) and is deliberately not offered as a mode: continuing a real prefix means shipping text welded to sentences the author actually wrote, and the same paper notes such outputs "may evade detection by stylometric means." That is the detector-evasion boundary, so the skill stops at few-shot.

Neighbors

Skill	Relationship
feynman	Plain-language translation; timbre is voice fidelity. Different goals
honest	Two-audience assessment writing; shares the "state your limits" discipline
consult	If the harness is installed, a genuine second model can play Role C — a real independence upgrade over role-switching
relay	The Producer/Reviewer handoff pattern this pipeline imitates in a single file
triangulate	The three-lenses-in-fixed-order shape, applied to engineering decisions
better-options	Run first if the brief itself may be wrong

Upgrade path. With the XYZ harness available, Roles A/B/C map onto real relay turns with separate model backends, which replaces role-switching with actual independence. This skill is the zero-install form of the same discipline, not a lesser version of it.

Provenance

Design constraints here are taken from two peer-reviewed/preprint sources, both verified against the arXiv record:

• Jemama, R. & Kumar, R. — "How Well Do LLMs Imitate Human Writing Style?" arXiv:2509.24930 · accepted, IEEE UEMCON 2025 · code: https://github.com/rajeshjnu2006/writing-style-uemcon2025 Supplies: exemplars over descriptions (few-shot up to 23.5× zero-shot; zero-shot summaries are not effective anchors); prompting strategy outweighs model size; fidelity and detectability are separable (human perplexity 29.5 vs machine ~15–16). Note: the abstract reports 15.2 and the results section 16.07 — an internal inconsistency in the source, immaterial to the design.
• Sawant, Y. G. — "PersonalBench: Measuring the Authorship Gap in LLM Personalization" arXiv:2608.19746 · code: https://github.com/yashsawant22/personalbench Supplies: the authorship gap (0.484–0.508 vs the 0.626 human floor); judge circularity — the LLM judge preferred a machine profile (0.542) over the real author (0.427); near-zero cross-metric correlation (|r| < 0.07), hence two firewalled lenses rather than one score.

Flagged for the reader: applied "humanizer" utilities — including tools circulated for papers and grant writing — are not used as a foundation here. They are applied products rather than evaluated method, they are contested within the research community, and their goal (reducing detector signal) is explicitly outside this skill's boundary. They are named only to mark the line.

Neither paper endorses this skill. The pipeline is an engineering response to their findings, and inherits their limits — most importantly that no amount of prompt-time work has been shown to reach human authorship.