solanabr/writer-style-skill/writer-style

Writer-Style

Write original content that sounds like a specific person, not generic AI. And, for technical topics, that is factually correct first and styled second. The engine is voice-agnostic; it ships with the `kaue` pack (Kaue / Superteam Brazil, tuned for Solana/Web3 education) as the default.

What this skill is for

Use it whenever the output should read like a named author:

• Educational Solana/Web3 content: courses, lessons, modules, explainers, tutorials, deep-dives, threads.
• Long-form technical writing in general, in a target voice.
• Drafting or editing/restyling existing copy into the voice.
• Building a new voice pack (add an author, build someone's voice, design a router) → the persona-builder agent.

Do NOT use for: brand-neutral/house-style copy; content that should sound generic; one-off "summarize this" where voice doesn't matter.

Honest scope. This produces content **in the author's register, with AI tells engineered out and facts

verified first**, not an indistinguishable clone. Voice fidelity is corpus-bounded and improves as more of

the author's on-register writing is added. (The default Kaue pack is calibrated from a small on-register

sample; lean on the exemplars and the naturalness floor, and don't over-promise "it's exactly him.")

The packs

An unqualified "write this in my voice" defaults to `kaue`. For another author/domain, load that pack instead; to create one, use the persona-builder. Everything below is the engine; the pack supplies the voices.

`david` routes differently. Its router (profiles/david/ROUTING.md) picks a register mode first,

then decides whether its one guest is on. That guest is `helius`, licensed for its evidence layer only —

its structural layer (numbered step lists, procedures) is banned, because importing it measured

−4.2 FK grades and −5 words of median sentence length against the corpus. Weight ≈ 85/15; primary-only

for non-technical pieces. The pack also inverts two engine defaults: enthusiasm is tail-loaded (cold

open, hot close, not hot edges), and its measured analogy rate is 0.0/1k, so its exemplar bank swaps the

analogy and verdict slots for worked-example and question-hinge. Its main container

(course-lesson) and anything code-bearing are extrapolations — the corpus has neither. Read

profiles/david/PACK.md before writing in it.

The two-layer model (summary — full detail in two-layer-model.md)

• Primary voice (profiles/kaue/kaue.md + kaue.card.yaml), the target author's own voice, always on,

reproduced (idiolect is the asset). It owns the tone, the seams, and the naturalness floor. Free, not counted toward the layer budget.

• Secondary voices (profiles/kaue/secondary/*.md + cards), transferable craft borrowed from master

writers, idiolect stripped (applied in Kaue's voice, never impersonation). One leads (backbone), at most one guests. Chosen by the piece's job, not its topic.

The inversion is the whole trick: reproduce the primary's idiolect, strip the secondaries'. Keep the

primary's quirks; borrow only the secondaries' craft.

The workflow — facts first, voice last (full detail in writing-workflow.md)

Persona prompting boosts voice but degrades factual accuracy — fatal for technical content. So voice is the last transform over verified facts, never the medium facts are discovered in. Five steps:

tools/validate_voice.py runs the Pass-C checks (tells, density, diff, audit); the voice-validator agent wraps them.

Running the bundled tools (any install). The validator and data live inside this skill's own directory,

so call them with absolute paths. Resolve the directory once and let `$SKILL` stand for it below

($CLAUDE_PLUGIN_ROOT/skills/writer-style in a plugin, .claude/skills/writer-style in an npm/project

install run from the project root, or skills/writer-style in a clone run from the repo root):

```bash

SKILL="${CLAUDE_PLUGIN_ROOT:+$CLAUDE_PLUGIN_ROOT/skills/writer-style}"

[ -d "$SKILL" ] || SKILL=".claude/skills/writer-style"; [ -d "$SKILL" ] || SKILL="skills/writer-style"

python3 "$SKILL/tools/validate_voice.py" tells --file <draft> --card "$SKILL/profiles/kaue/kaue.card.yaml"

```

Routing — which secondary voice (full router in profiles/kaue/ROUTING.md)

Route on the piece's dominant job, not its topic. Each voice is a full backbone in its lane:

Rules: primary always on & uncounted; cap = backbone (1) + guest (≤1) = 2 secondary layers (3 only for 3,000+ words, segregated; else split the piece); stack only across different lanes; helius is the safe universal guest; hold each voice's guardrail (Hayes builder-optimist not cynical · Balaji peer not oracular · Hotz inviting not dismissive · Vitalik warm not flat). The lanes self-differentiate: no author-vs-author special rules.

Explicit tone override. The requester may pin the blend — tone: hayes, tone: helius 0.7 / hotz 0.3, "a touch of vitalik", or primary-only. An explicit request outranks the router: apply it within the cap (backbone + ≤1 guest; rough weights set which is which). If the request crosses a lane guardrail (e.g. hotz on security), still honor it, but say so in the tone manifest instead of silently re-routing.

Tone manifest (always). Every delivery ends with ONE metadata line in the chat/report — never inside the piece itself: tone[helius 0.8, hayes 0.2] · mood: dry-competent · markers: sign-off 1 · route: auto Weights are the writer's honest estimate of each secondary's share of the craft (backbone-heavy by construction; the always-on primary is not weighted — it owns the floor). route: auto vs route: requested records whether the blend was router-chosen or user-pinned, so the requester can iterate ("same piece, hayes 0.5").

The naturalness floor (protect above all craft — rules/naturalness.md)

The body of every craft, applied thoroughly, smooths prose into something that sounds like nobody. Counter it: write unevenly: long calm stretches, enthusiasm only at the edges (open / seams / verdicts / close), and real human seams that CLUMP (a confession, a real number from your own use, a named credit, a "this ran long" — several per long piece, in floods and droughts, never one-per-section on schedule; a met quota is itself the tell). Even enthusiasm + even polish is the #1 AI tell. Always load the seam exemplar.

What you load at generation time (Pass B)

profiles/kaue/kaue.md (primary prose) · kaue.card.yaml (dials + markers: budgets/gates) · profiles/kaue/themes.md (the substance bank — pull FITTING stances/anecdotes instead of recycling exemplar content) · ~4 exemplars from exemplars/kaue/ (always the seam + opener + close + 1 body slot by job; exemplars are rhythm donors, never content donors) · the one routed secondary (secondary/<v>.md + card + its move-demos) · if the piece targets a specific container (X post/thread, LinkedIn, blog, course lesson, video script, newsletter, community announcement), the one matching contract in formats/ — container is orthogonal to job; it constrains the vessel, not the routing. Never read evidence/*.profile.json, calibration/, or evolution/ — builder/validator-only.

*Substitute the pack directory for another pack (profiles/david/david.md + card + themes.md + ~4 exemplars). For a primary-only pack there is no routed secondary to load — the router hands you a register mode instead.*