Beyond a forgotten fact
Early in a novel, a protagonist refuses to expose a friend's secret. Later, she volunteers it to a stranger because the plot needs a revelation. Her biography remains intact: job, age and appearance are correct. What slipped is the boundary that made her behavior recognizable, and nothing in the story explains the change.
A scene can preserve facts while altering motives, relationships or moral limits. Fluent language may conceal the substitution until another chapter rests on a version of the character you never intended. Characters can change, of course. The distinction is between a change the story earns and a convenient replacement that solves the current scene.
Give the story an authority
A useful canon separates established facts, character beliefs, provisional plans and unresolved questions. A suspicion belongs in a different category from the author's confirmed explanation. If suspicion becomes fact, a draft can reveal a mystery early or make later evidence redundant.
Record the source of important claims and when they become available to each character. The author may know who sent a letter long before the protagonist does. A scene brief should state the protagonist's knowledge at that moment, not provide an undifferentiated account of the whole book.
Motivation deserves the same care. She distrusts institutions is less useful than she will accept help from an individual clerk but will not sign a statement she cannot verify. The latter offers a behavioral boundary to test. Leave room for exceptions and identify what would make one believable.
An illustrative continuity problem
Consider an invented mystery, unrelated to published PennedCanon fiction. Iris's brother Leon is alive but missing. The municipal archive burned in 2019. Iris needs a brass decoding key to interpret their father's notes. These are hypothetical constraints, not findings from a model experiment.
A generic request for a tense archive search could put Iris in an intact reading room, let her decode a letter immediately or announce that Leon is dead. A scene might avoid all three errors and still make her trust the first official explanation without cause. Review must ask both what happened and whether this person would accept it on the evidence available.
A more precise brief locates the search in the burned basement, permits discovery of a coded fragment, withholds Leon's location and preserves the need for the key. Iris can cooperate with the archivist without trusting his account. The scene should end with a consequential choice about whom to approach for the key. This supplies a purpose while leaving space for invention.
The brief does not guarantee consistency. It makes requirements inspectable. A version that follows every rule but gives Iris nothing difficult to decide has preserved facts without necessarily producing a compelling scene.
Long context is not a guarantee
Liu and colleagues' 2024 research tested multi-document question answering and key-value retrieval. For the models and tasks studied, the position of relevant information affected performance, with difficulties using information in the middle of long inputs. This was not an experiment on fictional characters or a measurement of every current model. It supports caution about equating a large context window with reliable use of every detail. [1]
Retrieve the relevant record and verify the result. A concise scene packet can contain the immediate objective, preceding events, each participant's knowledge and the constraints most likely to be violated. Keep the fuller canon available. If packet and manuscript conflict, resolve the conflict instead of asking the assistant to choose silently.
Preserve changes deliberately
Distinguish an invention you reject from one you adopt. Accepted details should enter the canon with a source and a place in the timeline. Rejected details should not remain in a summary later prompts treat as authoritative. When someone changes their mind, record the event that caused it and the behavior it now permits.
Check consequences outward. If Iris learns the archivist's connection to Leon here, later conversations depending on her ignorance require review. Updating a biography cannot by itself repair a dependency running through multiple scenes.
Try a repeatable review
For an upcoming scene, list three facts, two beliefs, one behavioral boundary and one unresolved question. If comparing generic and structured prompts, retain model name, date, instructions and outputs. Judge both versions for contradictions, unsupported knowledge, motivation and dramatic movement. Do not call a single comparison proof that one method always wins.
Have a reader assess tension and clarity without knowing which brief produced which version. Where structure helps, identify the useful constraint. Where it produces stiffness, improve the conflict rather than discarding the canon. The author still decides which sentences belong; the record gives that decision something dependable to answer to.
Sources & editorial note
Prepared with AI research and drafting assistance. PennedCanon is the editorial publisher and has a commercial interest in the workspace. Examples are original illustrations, not recorded model experiments. Practical exercises are editorial recommendations, not guarantees of quality, authorship or copyright.