Check Consistency

Skill: check-consistency

What it does. Checks chapters against the project's established facts: world rules, timeline markers, character consistency rules, and knowledge states. Finds contradictions before readers do.

Triggers

• "check consistency of chapter X.Y"
• "/check-consistency {scope}"
• "does chapter X contradict anything?"
• "run a continuity check on Act {N}"
• Equivalent phrasing in the project's output language

Input. Scope: one chapter, an Act, or the whole book.

Output. Report in chat; confirmed contradictions logged to the contradictions table in .project/knowledge/worldbuilding.md.

Write the report in the project's output language.

Core principle

Flag only what is actually contradicted, with evidence from both sides. Never resolve silently.

A consistency check that cries wolf trains the author to ignore it. Every flag must cite two passages that cannot both be true — or one passage against one established rule. Ambiguity, stylistic variation, and things merely unstated are not contradictions.

And when a real contradiction is found, which side wins is an author decision. The skill presents; it never picks.

Execution protocol

Step 1 — Load the sources of truth

1. .project/config/project.yaml — output language, paths.layout
2. .project/knowledge/worldbuilding.md — rules (ESTABLISHED ones especially) and their costs
3. .project/knowledge/timeline.md — chronology and the time-markers table
4. .project/knowledge/characters/*.md — Consistency rules sections and knowledge states
5. .project/knowledge/glossary.md — canonical spellings and term usage
6. The chapter list in scope, in reading order — resolve via .project/templates/layout.md. Order is not optional here: timeline arithmetic and knowledge states ("who knows what, when") are only checkable against the sequence. An Act scope resolves from the chapter id prefix

Step 2 — Read the scope against them

Four sweeps:

a) World-rule sweep. Does anything in the text violate an established rule, ignore a stated cost, or exceed a demonstrated limit?

b) Time sweep. Do stated durations, sequences, and markers reconcile? ("three days later" chains are the classic breakage — sum them and check.)

c) Character sweep. Physical facts, behavioural invariants, voice register (flagrant breaks only — fine voice drift belongs to create-character's extract mode).

d) Knowledge sweep. The subtlest and most valuable: does any character use information they cannot yet have? Check reveals against who was present, who was told, and when.

Step 3 — Classify each finding

• CONTRADICTION — two citations that cannot both hold. Severity: does a plot point depend on it?
• STRAIN — not strictly contradictory, but requires a generous reading. Worth surfacing, marked as such.
• GAP — the text is silent where a fact is expected. Usually fine; listed only if a future chapter will need the answer.

Step 4 — Report

Per finding:

### {short label} — {CONTRADICTION | STRAIN | GAP}
**Side A:** "{quote}" ({chapter})
**Side B:** "{quote}" ({chapter}) — or: rule/sheet entry
**Stakes:** {what depends on this}
**Options:** {the 2-3 ways it could be resolved — no recommendation unless asked}

Step 5 — Log resolutions

When the author decides, record it in the worldbuilding contradictions table with date and resolution. If the resolution changes a rule or a character fact, update the relevant file (with the author's confirmation).

What to avoid

Pedantry. An unnamed weekday is not a gap. A metaphor is not a rule violation. The bar is would a careful reader trip here?

Silent resolution. Even an "obvious" fix changes the manuscript's commitments. Present, don't decide.

Re-litigating decisions. A contradiction the author already resolved (it's in the log) stays resolved.

Treating ASSUMED rules as binding. Only ESTABLISHED rules generate contradictions. An assumption in tension with the text is a signal to update the assumption, and is reported as such.

Cadence

After each Act, before any draft-scene session that touches long-established material, and always after a restructure-chapter plan is executed — moving scenes moves what the reader knows when, which is exactly what the knowledge sweep catches.

Relationship to other skills