Voice & Style
Overview
Record the book's voice and surface conventions in style-sheet.md, then
enforce them. The style sheet is the copyeditor's record: dialect and
spelling choices, capitalisation, hyphenation, numbers, dialogue
punctuation, a one-line voice summary per major speaker, and a watch list.
Its frontmatter is machine-readable, and story prose counts violations
and prose tics in every chapter. Use this skill so voice stays the same
across chapters, drafting sessions, and agents.
Prerequisites
A story project with story.md in the root. story init scaffolds
style-sheet.md; older projects can add it with story init '<title>' --dir . --force, which only adds missing files, or by copying
references/style-sheet-guide.md's frontmatter block by hand.
Language
Read language in story.md (a BCP 47 tag; a missing field means en).
The style sheet records the conventions of that language, and the voice
description is written about prose in it. dialect: british and
american and the built-in word lists behind filter words, -ly adverbs,
and said-bookisms are English. For another language, set
dialect: unspecified, record spelling variants as preferred entries,
and settle dialogue and punctuation with
../line-editing/references/language-conventions.md. When story prose
or story voices reports a check skipped for the book's language, do that
pass by reading (see references/prose-checks.md, Other languages).
When to Use
- Starting a project, once the first chapter or a writing sample exists
- Before drafting when voice has drifted between chapters or sessions
- When a copyedit pass needs a style decision recorded (the copyedit itself belongs to the
line-editing skill, which reads this file)
- When the user wants a mechanical prose check before sharing a draft
- Recording each character's voice (the style sheet's Character Voices line,
voice-words, voice-avoid) and checking it with story voices
- NOT for character personality or arc (use
character-management)
- NOT for scene-level craft such as deep POV or subtext (use
scene-craft)
- NOT for the line-by-line prose pass itself (line edit, copyedit, read-aloud, proof): use the
line-editing skill, which reads this style sheet and runs these checks
- NOT for rewriting dialogue so the voices sound distinct when everyone sounds the same: use the
line-editing skill's voice differentiation pass, which follows the voices recorded here
Workflow
1. Build or update the style sheet
- Read
story.md (genre, POV, tense, Tone & Style), style-sheet.md,
and two or three drafted chapters or a sample the user supplies. If
there is no prose yet, ask the user for a paragraph that sounds right or
three published books whose voice is close.
- Fill the sections in
style-sheet.md following
references/style-sheet-guide.md. Record decisions the prose already
makes; do not invent new ones silently. When the draft is inconsistent
(both grey and gray), ask the user which form wins.
- Set the frontmatter:
dialect: british | american | unspecified (unspecified for a
book not in English)
- one
preferred entry per variant: use is the house form, avoid
the form to flag. Repeat entries for several variants of one word.
watch-words: words or phrases this book overuses
allow-words: built-in filter words, adverbs, or dialect spellings
this book uses on purpose (a proper noun like Harbor Street in a
British book)
samples: files or folders of the author's own prose that sound right,
relative to the project folder: an earlier book (../book-one) or
chapters the author has approved, each file named on its own
(chapters/chapter-01.md, never chapters/, which warns
style-sample-own-chapters). A listed chapter is the measure, so
story prose does not compare it. Ask the user which; never list prose
the agent drafted and the author has not approved, or the profile
measures the agent's voice. With at least 2,000 words of narration,
story prose compares each chapter with the samples instead of fixed
limits
- Character Voices entries summarise each character file's Voice & Speech
Patterns section in one line and link to it. Record words a speaker
reaches for in the character file's
voice-words list and words they
would never say in voice-avoid, so story voices can check them. The character file is
canon; if the two disagree, fix the style sheet, or ask the user before
changing the character.
2. Draft and revise against it
chapter-writing and discovery-drafting read style-sheet.md before
drafting. Write in the recorded voice, use the house spellings, and give
each speaker the recorded patterns.
- When a new convention appears mid-draft (a new invented term, a number
style), add it to the style sheet in the same change.
3. Run the prose check
story prose .
The report lists, per chapter: sentence count, average and longest length,
and spread; filter words and -ly adverbs per 1,000 narration words
(dialogue is excluded); plain and said-bookism dialogue tags; words echoed
within 30 words; watch-word counts; and avoided spellings. Across the
manuscript it lists repeated 4-word phrases and character first names that
readers could confuse. Warnings are advisory. The command exits 0 unless a
file cannot be read, a story.md severity entry promotes a warning to an
error, or the command line is wrong, such as --baseline with no samples
(exit 2). See references/prose-checks.md for what each count means and how
to respond.
When the style sheet lists samples, the report opens with a profile of
the author's own prose, and chapters that drift from it warn
(prose-baseline-sentences, -paragraphs, -dialogue, -filter-words,
-adverbs) in place of the fixed filter-word and adverb limits. Treat a
drift as a prompt to reread the chapter, not a rule. Tell the user where
it drifts and ask whether it is deliberate: a fight scene runs short, a
quiet chapter long. Never rewrite a chapter just to move a number back
inside its tolerance.
To lint a passage that is not in a chapter file yet (a draft scene, a
proposed rewrite), pipe it in with - in place of the path:
story prose - < draft-scene.md. Run it from the project folder, or add
--path <project>, so the style sheet applies. story voices - does the
same for dialogue.
4. Run the dialogue voice check
story voices .
story voices fingerprints each character's attributed dialogue and
warns about a voice-avoid word said, a voice-words entry never said,
and two characters who may sound alike. A line counts only when the
narration names its speaker by a speech verb, or names one character in
the paragraph. Pronoun tags (she said) are never attributed, so in
close third person the POV character is often under-counted; when that
matters, name the tags in a sample chapter and rerun.
story voices --help describes the command and, when story-maintenance
is installed, the voices section of
../story-maintenance/references/continuity-checks.md
gives the full attribution rules and thresholds.
When two voices blur, differentiate them on more than one axis (sentence
length, contractions, vocabulary, what they ask about) and see
dialogue-subtext.md in the scene-craft skill for the tag-swap test.
Attribution and contraction rates are built around English quote marks,
speech verbs, and contractions. For a book in another language, if the
report attributes few or none of the lines the book has, or skips the
check for the language, compare the voices by reading: the levers
are the same (sentence length, vocabulary, forms of address, what each
avoids saying).
5. Act on the findings
- Fix avoided spellings everywhere; they are always errors of consistency.
- Treat rates, echoes, and repeated phrases as prompts to reread the
passage, not orders. Keep a flagged word when it is the right word, and
add it to
allow-words when the book uses it deliberately.
- For uniform-rhythm warnings, revise sentence length by intent (short for
impact, long for flow), not by formula.
- For similar names, ask the user before renaming; check replacements with
story names '<Candidate>'. Preview the rename with
story rename character <id> '<New Name>' --prose --dry-run, show the
user the replacements, then run it without --dry-run, so the chapter
text follows as well as the references. Without --prose the prose
keeps the old name. Check the matches it leaves as written as the
character-management skill's rename steps describe.
- For
story voices warnings, revise the dialogue or update the
character's voice-words/voice-avoid when the draft has found a
better voice; ask the user before changing canon.
- Present a summary: what changed, what was kept on purpose, and any
style-sheet updates. For a full line edit, copyedit, or proof, hand off
to the
line-editing skill.
CLI Maintenance
Use the Story CLI when it is available. If story is not installed, use the
bundled fallback node ../story-maintenance/scripts/story.js with the same
arguments. Use node <checkout>/bin/story.js instead only when the user
names a Story Skills repository checkout or you are working in one. Write
the script as an absolute path (resolve the fallback relative to this skill
folder) and run it from the folder you would run story from, so . and
other relative paths keep their meaning. Use Node, not Bun or a package
script: Bun would load that folder's bunfig.toml (which can run code) and
.env, and a package script runs from the checkout's root. If no CLI is
available, apply the checks in references/prose-checks.md by reading.
After editing the style sheet or revising prose:
story reindex .
story wordcount . --write
story check .
story prose .
story voices .
Reference Files
references/style-sheet-guide.md - What goes in each style-sheet section, the frontmatter format, and how to extract a voice description from sample prose
references/prose-checks.md - What each story prose count measures, its threshold, when to keep the flagged text, and what to do for a book not in English
../line-editing/references/language-conventions.md - Dialogue and punctuation conventions by language, for the style sheet's Dialogue And Punctuation section
Shared Conventions
Every story skill follows the shared conventions in ../story-maintenance/references/conventions.md, resolved relative to this skill folder. Read it before creating, renaming, or linking story files. If that file is missing because this skill was installed without story-maintenance, the essentials are: kebab-case ids and filenames, YAML frontmatter on every story-project file, _index.md registry tables that story reindex rebuilds (never edit them by hand), bidirectional links between entities, characters for who is on the page and mentions for who is only referred to, status: deceased plus died-in: chapter-{NN} for deaths, and no project-local generator or build scripts (run only the installed or bundled Story CLI).