Onboarding

Create the smallest useful first version of the writer's durable voice and style guides. Help them start; do not make them finish an exhaustive profile before they can write.

Load Existing Context First

Read ../../references/context-contract.md. Inspect the current workspace for maintained identity, voice, preference, and project-style files before asking questions.

• If authoritative context already exists, summarize it and ask only about genuine gaps.
• If the user needs a writing-home folder, resolve the target and run cw-setup-project first so VOICE.md, STYLE.md, examples/, and drafts/ exist together.
• Do not auto-trigger because either guide is sparse or because a legacy TASTE.md is absent.
• Do not duplicate global guidance into the writing home unless a local override is needed.
• Do not overwrite an existing profile without explicit approval.

Keep The Destinations Distinct

Route each rule by what it governs:

• VOICE.md answers How should the sentences sound? Put syntax, diction, and tone here, including cadence, rhythm, register, punctuation, verbal tics, sentence moves, and phrases to favor or avoid.

• STYLE.md answers What must the article do, contain, and prove? Put argument, evidence, article structure, substantive standards, audience promise, openings, endings, formats, and publication-readiness criteria here.

• Optional AUDIENCE.md answers Who are we writing for, and what may engage them? Put reader situations, prior knowledge, wants, interests, resistance, and source evidence here. Follow a maintained shared source named by the project; do not create a local duplicate.

Use this classification test:

• A rule that changes wording, sentence construction, or tone belongs in VOICE.md.
• A rule that changes the claim, support, organization, or readiness standard of the article belongs in STYLE.md.
• Reader knowledge belongs in the governing AUDIENCE.md; the writing's obligations to that reader remain in STYLE.md. Split mixed feedback without duplicating it across guides.

Use the writing home's VOICE.md and STYLE.md as its portable context. Use an existing global voice guide when it is already the maintained authority. Tell the user where each class of guidance will live when the destination is not already obvious.

Begin Or Refresh Audience Context

When the user requests audience setup or calibration, resolve the audience source through the context contract. Use an existing shared guide when named; otherwise create AUDIENCE.md in the established writing home if useful. The optional starter is defaults/project-template/AUDIENCE.md relative to the plugin root. Missing audience context never blocks the immediate writing task or triggers onboarding by itself.

Start with known situations, knowledge, interests, and resistance. Ask only consequential gaps. Label reader evidence, positioning, explicit writer intent, and editorial hypotheses; keep source, date, and scope. Do not invent a demographic profile, assume personal stakes, or turn a publication's intended relationship into measured reader behavior. The assignment can select a narrower reader without changing the shared guide.

Make the guide legible to both the writer and the agent: begin with a short reader portrait, distinguish wants (outcomes or experiences sought) from needs (what may help the reader get there), and connect them to what the writing can offer. Include a possible hero's journey when useful: starting situation, call, obstacles, guidance, trials, and what the reader carries back. Label it as an editorial model unless supported by reader evidence; do not force a transformation story or invent motives to fill its stages. Close with concise instructions for applying the guide to a piece, with sources and uncertainties easy to find.

Interview

Ask only what the existing context does not answer. During first-run setup, ask one to three conversational questions at a time or invite the writer to share representative samples. Useful dimensions include:

• syntax, diction, tone, sentence architecture, rhythm, and recurring line-level moves;
• the writer's argument standards and recurring substantive moves;
• evidence requirements, article structures, openings, endings, and publication-readiness criteria;
• what the writer makes, for whom, and with what reader promise;
• verbal tics, structural anti-patterns, and AI tells to avoid;
• workflow preferences that materially affect the project.

Ask conversationally and follow the user's energy. Six fixed questions are not a requirement.

For a quick first version, capture a few high-signal rules in each guide and leave unknowns open. If the writer arrived with an immediate task, calibrate from that work where possible instead of pausing for a separate interview.

Calibrating A Voice Profile From Examples

When the writer supplies a fresh sample set or asks to refresh VOICE.md, make the deliverable a sentence-level vocal score, not a writing philosophy.

First, compare the examples for repeatable line-level behavior. Capture only patterns that recur across the sample set or that the writer explicitly confirms. Separate patterns that belong to a particular publication, assignment, or subject from portable voice guidance.

The resulting VOICE.md should normally cover:

• the overall sentence-level sound in a few concrete sentences;
• cadence and sentence-length variation;
• syntax and recurring sentence turns, such as accumulation, qualification, correction, release, or fragments;
• diction and register, including words, constructions, and generic phrasing to avoid;
• punctuation, emphasis, and paragraph-level texture only when they affect sentence sound;
• humor, self-awareness, or verbal tics when they are evidenced in the samples;
• practical revision instructions that help an agent align a draft without imitating the examples.

Do not put curiosity, intellectual rigor, uncertainty, point of view, reader relationship, argument, evidence, structure, or publication standards in this profile unless the instruction changes the wording, syntax, cadence, or tone of a sentence. Route article standards to STYLE.md or the brief, and durable reader knowledge to the resolved AUDIENCE.md when one is maintained.

For every material voice rule, attach either a short sourced excerpt or a concise description of the observed move. Full examples belong in examples/; published language is directional and non-reusable. Avoid broad labels such as “warm,” “smart,” or “authentic” unless they are translated into observable writing behavior.

Write The Profile

• Capture only specific, evidenced preferences.
• Keep syntax, diction, and tone in VOICE.md; keep argument, evidence, article structure, substantive standards, and publication readiness in STYLE.md.
• Phrase guidance as actionable rules, not personality adjectives alone.
• Put full exemplars in examples/; use short excerpts in the guides only when they clarify a rule.
• Mark published examples as directional and non-reusable.
• When writing from examples, make rules operational: tell the agent how a sentence gathers, turns, qualifies, lands, or sounds false.
• Include a concise revision checklist for preserving the writer's sentence-level sound without copying their phrases or inventing personal detail.
• Show the proposed content before changing a high-authority or shared file when the governing rules require approval.

Migrate TASTE.md

When a legacy TASTE.md exists, classify its rules into VOICE.md, STYLE.md, and an authorized audience guide when useful. Flag mixed or conflicting rules for judgment. Do not keep adding new guidance to TASTE.md, and do not delete it without explicit approval.

Handoff

Confirm:

• what was captured in voice, style, and audience context when requested;
• where the guides and supporting examples were saved, including any shared audience source;
• which gaps remain intentionally open;
• how future Compound Writing skills will load the writing home.