Mine a writing corpus for an author's recurring stylistic habits — sentence rhythm, paragraph shape, transition words, syntactic moves, vocabulary register, hedging patterns — and turn the findings into a concrete DO/DON'T checklist that future drafts can imitate or avoid.
Give this skill a handful of writing samples from one person (or one consistent voice — a brand, a publication, a character), and it walks through a measured, evidence-first procedure to answer one question: how does this writing actually sound, specifically, and how would you reproduce that on purpose?
It doesn't guess at vibes ("sounds formal," "sounds casual"). It counts things — sentence lengths, transition-word frequency, semicolon rates, hedging phrases, first-person pronoun use — cross-checks the counts against close reading, and only then converts what it finds into imperative rules a writer or drafting agent can actually apply: use word X as your default pivot, not "however"; never close a section on a bullet recap; scope-narrow your hedges instead of softening them.
It also includes a specific, useful sub-technique: telling genuine, deliberate hedging apart from generic AI-sounding throat-clearing ("it is worth noting that," "arguably," "to some extent"), so an editing pass aimed at removing AI-texture doesn't accidentally erase a writer's real voice along with it.
One hard rule the whole method is built around: the output profiles voice, never content. It never extracts arguments, facts, or reusable material from the source corpus — only the shape of how that corpus is written. Any future draft that wants to reuse an idea from the corpus needs separate sourcing; the style profile is not a shortcut for that.
This is a portable Agent Skill — a folder following the SKILL.md convention, not tied to
any one tool.
- Claude Code: copy this folder into
.claude/skills/author-style-profile-mining/in your project (or the equivalent user-level skills directory), and it will be available the next time you start a session. - Any other agent runtime that supports the
SKILL.mdconvention: copy this folder into that runtime's skills directory following the same layout.
No dependencies to install and no script to run — the whole skill is the SKILL.md
instructions plus the reference template.
examples/synthetic-corpus/ contains three short, entirely fabricated blog-style posts from a
made-up voice ("Backyard Beds," a fictional gardening blog written for this example only).
examples/synthetic-style-profile.md is the full profile produced by running this skill's
procedure against those three files — computed from real counts against the sample text, so you
can see exactly what a completed profile looks like at every phase: the sentence-length
distribution table, the transition-marker table, the hedging-vs-throat-clearing classification,
and the final DO/DON'T list.
A short excerpt of the pattern it surfaces:
DO: Open by stating what the author believed and acted on, then pivot with "Turns out" to the corrected finding — not "however" or "it turns out that," the bare two-word phrase.
DON'T: Don't scatter soft qualifiers ("arguably," "to some extent") through ordinary declarative sentences — this author's hedges are concentrated, named, and scope-narrowing only.
Every rule in that list traces back to a specific count or quoted instance in the profile — that traceability is the point of the method, and worth preserving in any profile you build from real material.
references/output-template.md is a ready-to-fill markdown skeleton mirroring every phase in
SKILL.md — copy it as the starting point for a new profile rather than building the section
structure from scratch each time.
MIT. See LICENSE.