Continuity editor

Read the book against its own canon and timeline and flag the contradictions — before a reader finds them.

The one rule

This skill reads, cross-checks, and flags. It never writes or rewrites the author's prose, and it never decides which version is canon. It surfaces the contradiction and both offending passages; it does not pick the "correct" fact or patch the text. Which version holds is the author's call — the skill only makes the collision impossible to miss.

Craft criteria

It reasons from continuity-editing practice, grounded in the book's own facts (declared canon when the author provides it, inferred canon otherwise). It tracks:

• Character facts — names, appearance, age, relationships, established backstory.
• Timeline — event order, elapsed time, days of the week, seasons, character ages across the span.
• World rules — the story's own established constraints (magic costs, technology limits, geography) and violations of them.
• Object & state continuity — a wound that heals too fast, an item held then never picked up, a locked door later open.
• Knowledge state — who knows what, and when; a character acting on information they haven't yet learned.
• Spatial logic — travel times and distances that don't reconcile.

Scope

Set at intake, because a continuity pass on a long book can't be held in memory all at once:

• Full audit — build a canon model of the whole book and cross-check every thread. On a long manuscript this is done in systematic passes, and the report states what was covered and what wasn't.
• Targeted check — verify one thread the author names: a single character's facts, one object, the timeline alone, or a specific worry ("does the wound in chapter 9 track?"). Fast, and context-friendly.

Input

Works on whatever the writer provides — pasted text, an attached file, or a named chapter or range. No tool or account assumed. It builds a working model of canon from the text itself; the more of the book it sees, the more contradictions it can catch. Standalone, it flags what's inferable from the provided pages, and says what it couldn't see.

Try it

You don't need to track it yourself — that's the whole point. Openers that work:

• "Here are my chapters — does anything contradict itself?"
• "I think a character's eye colour changes somewhere. Can you track it?"
• "Some gaps are deliberate mysteries, not mistakes — flag them as questions, not errors." (declare intent)

Intake — what it settles first

Before checking, settle these — ask only what the submission doesn't answer:

1. Scope — full audit or a targeted check (above).
2. Declared canon — any story bible, character sheet, series facts, or timeline the author wants treated as fixed truth. Optional; absent it, canon is inferred from the manuscript, and the skill says so.
3. Intentional devices — what's deliberate: an unreliable narrator, characters who lie, withheld secrets and deferred reveals, multiple or non-linear timelines. These are pre-cleared so a designed effect isn't reported as an error. Name the load-bearing ones — declaring everything intentional defeats the check.
4. Known-fixed facts — anything the author is certain of and wants held as the reference.

The method

Standalone, the skill has no canon database — it builds one, the way a fiction copyeditor builds a style sheet (the trade's canon artifact — Amy Schneider's Chicago Guide to Copyediting Fiction; CIEP's fiction style-sheet practice). It extracts a compact record organized like that sheet:

• Entities — a fielded entry per character and place: appearance, age, relationships, established facts. Where an age and a date both appear, it does the math (a character "14 at the end of Book One, born May 2009" is a checkable pair).
• First occurrence — every fact records where it was first stated, so a later collision traces straight back to the page it contradicts.
• Timeline — a running chronology rebuilt from temporal cues: event order, elapsed time, day-of-week and season, ages across the span.
• Confidence per fact — a fact the text states is firmer than one the skill infers; the record marks which, because a flag is only as strong as the weaker of the two facts it collides.

Error vs. device — the heart of the skill. Every collision is reported as a question about intent, never an assertion of error — the professional copyeditor's query posture (CIEP; Chicago Manual of Style; SFWA), because the skill cannot know which version the author intends. Two blue-eyed parents with a brown-eyed child is apparently impossible — or a deliberate hint at an affair; the skill asks, it does not correct. A collision drops from confirmed to suspected when a pre-cleared device could explain it — an unreliable narrator, a character who might be lying, a secret not yet revealed, a deliberate retcon.

One limit is built in honestly: telling a genuine slip from a deliberate cross-narrator contradiction or a withheld-information reveal is hard — the current evidence is that even careful automated judgment is weak at it. So those calls are always surfaced low-confidence, and for deferred reveals the skill reasons in two modes — what the reader knows before the reveal versus after — rather than flagging a planted secret as a mistake.

How it runs

1. Intake — scope, declared canon, intentional devices, known-fixed facts.
2. Extract canon — build the working record from declared canon and the text; note confidence per fact.
3. Cross-check — read each scene against the record for facts, chronology, world rules, knowledge-state, and spatial logic.
4. Classify — each collision as a confirmed contradiction or a suspected one (naming the device it might be); mark confidence.
5. Report — grouped by type, both passages cited, resolution left to the author.

Output

A markdown continuity report — every entry a query about intent, never a correction, because which version is canon is the author's call. Two tiers keep a designed effect from masquerading as a mistake:

• Confirmed contradictions — the text genuinely says both things (chapter 1 versus chapter 12) and no pre-cleared device explains it. Confirmed means the collision is real, not that it's a mistake — the author still decides which fact holds.
• Suspected (possible device) — a collision a device might explain, named for what it could deliberately be, for the author to confirm or dismiss in seconds. Cross-narrator and withheld-secret cases sit here by default, at low confidence.

Within each tier, flags are grouped by type (character, timeline, world, state, knowledge, spatial). Every flag cites both conflicting passages with locations and the point each fact was first established, states the conflict plainly, and carries a confidence mark. Where the timeline is in question, it includes the extracted timeline the author can check against their intent. No resolution is written into the prose; the author chooses which fact holds.

The shelf

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

• The style sheet. Amy Schneider, The Chicago Guide to Copyediting Fiction (its Characters / Places / Timeline structure), and CIEP's fiction style-sheet practice — the trade's model for a compact canon record.
• The query posture. The Chicago Manual of Style and CIEP on raising an author query rather than silently changing, and the SFWA copyeditor's guide (Terry McGarry) on catching errors of logic and continuity — and its catalogue of the classic slips.
• Declared canon. The author's story bible — the novelist's adaptation of television's show bible — as the consolidated facts a manuscript is checked against.

With Calliope (MCP)

Connected over the MCP connector, it reads the real book — fetching chapters and searching every mention of a fact across the manuscript — and reads the Map as declared canon, so it checks the prose against stated truth instead of re-inferring it. Because this skill writes back, with a subscription it files each flag as marginalia at the exact conflicting scenes and can record a reconciled timeline note in the Map. It still never rewrites prose and never picks canon — continuity flags and notes only.