Phase 0: Onboarding

Goal: enough shared understanding that every later agent writes the same book the author imagines. The two documents produced here, PRD.md and STYLE.md, are read by every agent on every task. Vague answers here become vague chapters later.

Read ${CLAUDE_PLUGIN_ROOT}/skills/loom/quality-bar.md first.

Interview

At most 3 questions per round, at most 4 rounds. Ask only what cannot be inferred from what the author already said or from materials they pointed to. If the author gives a rich brief up front, skip straight to confirmation.

Round 1 — What and for whom

• Genre and subgenre (fiction/nonfiction/technical/screenplay/poetry/game/academic; for fiction, the shelf it sits on: 스릴러, 로맨스, literary, SF, etc.)
• The reader: one or two concrete people, not demographics. "A 34-year-old backend engineer who has shipped LLM features and been burned by them" beats "developers."
• Language, and if Korean or Japanese, register conventions (존댓말 narration? 문어체?).

Round 2 — Scale and materials

• Length. Offer the genre norm from quality-bar.md language notes; let the author adjust.
• Existing material: notes, outlines, prior drafts, research, code, an Obsidian or alcove vault. Where does it live?
• Timeline and any hard constraints (series continuity, publisher spec, trim size).

Round 3 — Voice

• Ask for a voice sample: 500-2,000 words of the author's own prose in the register they want, or 2-3 published authors whose voice is the target, or both. If neither exists, offer to write three 200-word openings in different voices during ideation and let them choose.
• Tone words the author would use and words they would reject.
• Anything the author never wants to see (patterns, words, tropes, structures).

Round 4 — The promise (can merge into Round 1 for experienced authors)

• What does the reader have at the end that they did not have at the start? One sentence. This becomes the north star for beta-reader.

Voice fingerprint

From the voice sample or named authors, write a Voice Fingerprint section in STYLE.md. This is the single most important input to chapter-writer. Be specific and quantitative where possible:

• Sentence length: typical range, how short the shortest go, how often a long sentence appears, what a long sentence is for in this voice
• Paragraph shape: one-liners common or rare; longest paragraphs do what
• Diction register: plain/literary/technical; Latinate vs. Germanic (English); 한자어 vs 고유어 balance (Korean); slang, profanity, jargon policy
• Figurative density: metaphors per page; what domains they draw from; what the voice never compares things to
• Narrative distance (fiction): how close to the POV character's thoughts; free indirect style or not
• Humor: kind, frequency, targets
• Dialogue conventions: tags (said only? none?), beats, dialect, 존댓말 map for Korean
• Signature moves: two or three things this voice does that another would not
• Three short quoted examples from the sample (or reconstructed exemplars if the author named published writers; never quote living authors at length)

If the author supplied a sample, also store it verbatim under ## Voice Sample in STYLE.md.

Source indexing

If the author has materials, scan them now (find, alcove search_project_docs/search_vault if available, Obsidian tags). Write sources/INDEX.md:

# Sources

| ID | Path or URL | Type | What it contains | Trust | Use for |
|----|-------------|------|------------------|-------|---------|
| S01 | sources/interview-kim.md | interview transcript | ... | primary | ch03, ch07 |

Trust levels: primary (author's own data, interviews, code they wrote), secondary (published, citable), tertiary (blogs, forums; use for leads only). Every factual claim in nonfiction drafts must cite an ID from this index or a verified URL. Copy or symlink materials into sources/ when practical.

Outputs

PRD.md — YAML frontmatter (title, genre, subgenre, language, target_length, count_unit, chapters) then sections: Identity, Reader (personas), Reader Promise, Scope, Constraints, Source Map (summary of INDEX), Success Criteria (measurable, including "readiness PASS from beta-reader"), Open Questions.

STYLE.md — Voice Fingerprint, Voice Sample (if any), Language Rules, Formatting Rules, Prohibited (author's list + quality-bar.md taxonomy reference), Dialogue Conventions, Genre Conventions (pointer to genre skill with any overrides). Leave a ## Voice Lock heading with "pending" so the router knows the lock has not happened.

Directories — drafts/ edits/ publish/ sources/ .velith/. Offer git init if the directory is not a repository; the pipeline snapshots drafts before rewriting stages regardless.

Gate

PRD.md and STYLE.md exist. STYLE.md has a Voice Fingerprint with at least five specific attributes. PRD.md has a one-sentence reader promise. Source index exists if the author has materials.

Post-completion

node ${CLAUDE_PLUGIN_ROOT}/velith.mjs scan [project-dir]