Workflows and subagents for consistency checks
Most of the individual checks described elsewhere on this site (canon consistency, farming facts, the language-police pass) are defined as Claude Code “skills,” reusable instruction sets a session can invoke on demand. Run one at a time on the chapter currently being drafted, a skill is just a normal part of an interactive session. Checking a whole book’s consistency in one sweep is a different job. The same checks now have to run across every accepted chapter at once, and that needs something more structured: a scripted workflow that fans work out across many subagents and collects what comes back, with nobody watching each individual check happen.
Why this needed its own workflow
A single session, even with a skill loaded, checks one thing against a manageable amount of context at a time. Asking whether one chapter is consistent with everything established before it is a check of that shape. Asking whether a whole book of them is internally consistent is that same check, run dozens of times, against a canon (the site’s term for the book’s accumulated, established facts) that keeps changing as earlier findings get resolved. That’s a different kind of task. It’s less about depth of judgement on any one chapter and more about coverage across all of them.
The site’s author, David, asked directly for this shape of tool partway through the project: a workflow that runs a bounded, cheaper model against each chapter individually, hunting for inconsistencies, rather than a single expensive pass trying to hold the whole book in mind at once. The reasoning is the same tradeoff behind the language-police pass. A smaller model doing narrow, well-specified work on many chapters at once, followed by a synthesis step reconciling what comes back, is usually cheaper and more thorough than one large context window trying to do it all in one pass. The session where this was specified and built is the clearest record of the reasoning behind it: 2026-08-27, building a consistency-check workflow.
How the pieces fit together
The consistency-check skill itself defines the actual judgement: read a chapter, check its claims against canon in priority order (hard facts and continuity first, softer thematic consistency last), and report contradictions, or the weaker “drift” category where a claim is technically compatible with canon but strains it under the weight of what’s already been established. A workflow wrapping this skill assigns one subagent per chapter, running these checks in parallel rather than one after another, and then a further step gathers every subagent’s findings into a single report rather than leaving David to read through dozens of separate outputs.
This division of labour, a shallow parallel pass followed by a deeper serial synthesis, shows up more than once in this project. It shows up in the language-police pass, where a fan-out of lens agents feeds one aggregating judge. It shows up in the per-scene editing pipeline, where four parallel lenses feed one synthesiser. And it shows up here, in per-chapter consistency subagents feeding one workflow-level report. In each case the parallel step stays cheap and narrow, and the serial step is where the actual judgement happens, including the judgement about which findings are even worth keeping.