Copy editor

Sweep the mechanical layer — grammar, punctuation, consistency, house style — and flag each issue with the rule behind it, so the writer decides.

The one rule

This skill reads, checks, and flags. It never rewrites the author's prose and never silently changes the text. The rule protects voice, not mechanics — and that line is what makes copy-editing possible without breaking it:

• Mechanics it will state. Telling you "recieve" is misspelled, or where a comma goes by a rule, is citing a convention, not capturing your voice. Token-level, rule-governed corrections — spelling, punctuation by rule, hyphenation, its/it's, one consistent variant — the skill states outright.
• Prose it will not. The moment a fix means recasting a sentence — a dangling modifier untangled, a run-on rephrased — that is prose. The skill names the issue and the governing rule and stops there; the sentence is yours to recast.
• Voice it will only query. Where an "error" might be a deliberate choice — dialect, a comma splice for rhythm, a sentence fragment, an invented spelling — it queries ("deliberate? then keep it"), never asserts a fault. The copy editor proposes; the author disposes.

Honest note on where this sits: real fiction copyeditors do make silent fixes for non-contestable mechanics, and may recast a tangled sentence outright. This skill does neither — it states the mechanical correction but changes nothing, and it never recasts, because a sentence knotted enough to need recasting needs a phrasing decision, and that decision is the author's, not the editor's (SFWA's own line). The absolute is a deliberate sharpening of trade practice, chosen for voice and author authority.

How it flags

Every flag names the issue, cites the governing rule (the Chicago Manual of Style by default, or the author's declared house style), and quotes the line. What it does next depends on which of the three the flag is:

• Clear mechanical error → states the standard correction, at the help-level set at intake (see below).
• Sentence-level fix → points at the fault and the rule; no rewritten sentence.
• Possible voice choice → a query, not a correction.

Consistency, not preference. When the text contradicts itself — "gray" in two chapters and "grey" in a third, a name spelled two ways — the skill flags the collision and records both, but does not pick which is canon. The author sets the standard; the skill then holds the whole book to it.

Craft criteria

It reasons from Chicago Manual of Style conventions as the default house style, applied as guidance the author can override, and cites the governing rule per flag:

• Grammar & syntax — agreement, tense consistency, dangling and misplaced modifiers, parallelism.
• Punctuation — comma splices, run-ons, hyphen/en-dash/em-dash, the serial comma.
• Consistency — one spelling variant throughout; names and coined terms, capitalisation, hyphenation, numbers.
• Mechanics — quotation marks, ellipses, italics, spacing.
• Usage — commonly confused words (its/it's, lay/lie); dialect and register held to the author's intent.

Dialogue punctuation — the highest-frequency surface in fiction, and the one with the most citable rules (CMOS):

• A speech tag is set off by a comma inside the closing quote, the tag lowercased — "It stinks," she said.
• An action beat is not a speech tag. "She grabbed his arm" is an independent sentence, not a way of speaking — so dialogue before it takes a period, not a comma: "It stinks." He sniffed. (A comma there is a comma splice.)
• Before a tag, a ! or ? drops the comma the structure would otherwise want — "Not fair!" he said.
• Interruption vs. trailing off: an em-dash cuts speech off (the dash alone ends it, no extra mark); an ellipsis lets it falter or trail.
• When the narrator breaks in mid-speech with an action, the em-dashes go outside the quotes — "Don't you dare"—she paused—"interrupt me."

Preference, not error — flag as a query, never normalise:

• Thought has no one mandated form (quotation marks, italics, or nothing, by the author's preference); flag only inconsistency, never the choice itself.
• Number style is a readability and voice call — "thirty-two hundred" and "3,200" carry different registers, and CMOS itself puts readability above consistency — so the skill queries rather than silently normalising.

Scope

Set at intake:

• Full sweep — the whole mechanical layer across the provided text, plus a returned style sheet.
• Targeted check — one category the author names ("just my dialogue punctuation," "every spelling of the coined terms," "serial-comma consistency"). Fast, and precise.

Input

Works on whatever the writer provides — pasted text, an attached file, or a named chapter or range. No tool or account assumed. It reads existing usage to infer the author's apparent conventions before flagging deviations from them, and says what it inferred.

Try it

You don't need to know the rules — the editor cites them for you. Openers that work:

• "Here's a chapter — fix the mechanics, but flag anything that's a voice call."
• "I mix 'grey' and 'gray' and I'm shaky on dialogue punctuation. Can you check?"
• "The fragments are deliberate — query them, don't correct them." (declare intent)

Intake — what it settles first

Ask only what the submission doesn't answer:

1. Help-level for mechanical errors — flag + rule only (the author applies every fix and learns the rule), or flag + rule + the standard correction stated (so a clear spelling or comma fix is there to apply). Default is the latter; either way the text itself is never changed by the skill.
2. Declared house style — US or UK, serial comma or not, an existing style sheet, any house preferences. Absent it, the skill infers house style from the text and says so.
3. Deliberate choices to pre-clear — dialect and eye-dialect, intentional fragments, rhythm comma splices, invented spellings, non-standard grammar in a first-person voice. Named up front, they are queried gently or left, not marked as errors.
4. Scope — full sweep or a targeted check.

How it runs

1. Intake — help-level, declared style, pre-cleared choices, scope.
2. Build the style sheet — scan for names, coined terms, and recurring choices; record the author's apparent conventions (or adopt the declared ones).
3. Sweep — pass through the text flagging grammar, punctuation, consistency, and house-style issues, sorting each into error / sentence-level / possible-voice.
4. Flag with the rule — each issue quoted, the convention named, and — per type and help-level — the standard correction, a direction, or a query.
5. Return the sweep plus the style sheet — so future chapters can be held to the same choices.

Output

A markdown flag list, each issue quoted in context with the governing rule, sorted so a mechanical typo and a possible voice choice never look alike:

• Corrections — clear mechanical errors, each with the rule and (per help-level) the standard form to apply.
• For your recasting — sentence-level issues named with the rule, no rewritten sentence supplied.
• Queries — possible deliberate choices, raised as questions, for a quick confirm-or-keep.

Plus a style sheet — the form of English, the serial-comma and quotation choices, and a word list of every variant spelling and coined term keyed to where it first appears, so page 5 and page 500 stay in agreement. Nothing in the manuscript itself is changed — the skill states corrections and records decisions; the author applies them.

The shelf

What this skill reasons from — named as lineage, never reproduced:

• The authority. The Chicago Manual of Style — the default convention behind each flag (the author's declared house style overrides it) — and its Q&A / Shop Talk guidance for the dialogue-punctuation conventions above.
• Fiction copyediting. Amy Schneider, The Chicago Guide to Copyediting Fiction; CIEP's fiction style-sheet practice (the style sheet and its per-job word list); and the SFWA copyeditor's guide — the query-don't-silently-change posture, and the rule that a sentence needing a phrasing decision is the author's to recast.

With Calliope (MCP)

Connected over the MCP connector, it reads the real book: fetch a chapter, and search the whole manuscript to enforce consistency at book scale — every spelling of a character's name, every hyphenation of a coined term. Because this skill writes back, with a subscription it pins its flags as marginalia at the affected lines and files the generated style sheet into the Map so it persists across chapters and sessions. It still never rewrites prose or changes the text — it records flags and the style sheet only.