Problem
The AI-assisted commands assume a novel.
OUTLINE_SYSTEM (packages/bookkit/src/bookkit/prompts/outline.py) casts the model as a developmental editor working on a "story spine", and OUTLINE_USER demands a Premise, a Voice & Style section, and a YAML canon of characters[] with role: protagonist | antagonist, voice, arc, status: alive. Point it at "a plain-language guide to CI/CD" or "a monograph on catalytic reduction" and it will invent characters — because that is what the schema it was handed asks for.
CHAPTER_SYSTEM is milder but still carries five continuity rules about characters, dialogue and established names, and none about the things that matter elsewhere: never using a term the reader has not been given, never overstating what a source supports.
Proposal
Three prompt families, selected by the bible's kind (#9) rather than a flag the author must remember on every invocation — with --kind available to override for a one-off.
technical
Outline — a developmental editor for explanatory non-fiction, emitting an Argument rather than a premise (what the book claims, what the reader can do at the end), a Reader section (who this is for and, critically, what is assumed known — the thing outlines for beginners always get wrong), and per chapter the objective plus concepts introduced versus assumed. Structured canon is a concept ledger, not a cast list.
Chapter — replace the character rules with the ones that apply: never use a term before its defined_in; define each term once, in the words the ledger commits to, keeping the committed analogy; do not silently escalate assumed knowledge; concrete example before general rule.
scholarly
As technical, plus: bind each symbol to one quantity throughout; do not assert a claim the provided sources don't support; hedge to the strength of the evidence rather than the strength of the argument. This family should be the most conservative of the three — a model that invents a plausible citation is worse than useless in a thesis, so the prompt must forbid producing citation keys that are not in the supplied bibliography (#15), and #10's citation-missing rule is the backstop that catches it when it does anyway.
Recap
For non-narrative genres, recap --chapter N --apply should record which concepts a chapter actually defined and which it used, feeding the ledger the way state_changes feeds character status today. That turns the recap from a plot summary into the mechanism that keeps defined_in honest — and gives #10's rules real data instead of the author's intent.
Scope
Depends on #9. Part of the genre generalization tracked in the umbrella issue.
Problem
The AI-assisted commands assume a novel.
OUTLINE_SYSTEM(packages/bookkit/src/bookkit/prompts/outline.py) casts the model as a developmental editor working on a "story spine", andOUTLINE_USERdemands a Premise, a Voice & Style section, and a YAML canon ofcharacters[]withrole: protagonist | antagonist,voice,arc,status: alive. Point it at "a plain-language guide to CI/CD" or "a monograph on catalytic reduction" and it will invent characters — because that is what the schema it was handed asks for.CHAPTER_SYSTEMis milder but still carries five continuity rules about characters, dialogue and established names, and none about the things that matter elsewhere: never using a term the reader has not been given, never overstating what a source supports.Proposal
Three prompt families, selected by the bible's
kind(#9) rather than a flag the author must remember on every invocation — with--kindavailable to override for a one-off.technicalOutline — a developmental editor for explanatory non-fiction, emitting an Argument rather than a premise (what the book claims, what the reader can do at the end), a Reader section (who this is for and, critically, what is assumed known — the thing outlines for beginners always get wrong), and per chapter the objective plus concepts introduced versus assumed. Structured canon is a concept ledger, not a cast list.
Chapter — replace the character rules with the ones that apply: never use a term before its
defined_in; define each term once, in the words the ledger commits to, keeping the committed analogy; do not silently escalate assumed knowledge; concrete example before general rule.scholarlyAs
technical, plus: bind each symbol to one quantity throughout; do not assert a claim the provided sources don't support; hedge to the strength of the evidence rather than the strength of the argument. This family should be the most conservative of the three — a model that invents a plausible citation is worse than useless in a thesis, so the prompt must forbid producing citation keys that are not in the supplied bibliography (#15), and #10'scitation-missingrule is the backstop that catches it when it does anyway.Recap
For non-narrative genres,
recap --chapter N --applyshould record which concepts a chapter actually defined and which it used, feeding the ledger the waystate_changesfeeds character status today. That turns the recap from a plot summary into the mechanism that keepsdefined_inhonest — and gives #10's rules real data instead of the author's intent.Scope
technicalandscholarlyoutline / chapter / recap prompt modules, dispatched onkind--kindoverride on thewritecommandsdocs/continuity.md§10Depends on #9. Part of the genre generalization tracked in the umbrella issue.