Give AI Scene Rewrites a Knowledge Ledger

Disclosure: I’m building Laper. This article was generated with AI assistance. The example is fictional, and the data structure below is a proposed workflow, not a claim about a feature in Laper.

A rewrite can preserve every character name and still break a mystery. The detective asks about a hidden recording before anybody has told her it exists. The line sounds natural. The scene no longer makes sense.

For an AI-assisted rewriting workflow, “keep the story consistent” is too broad a check. One smaller question is easier to review: what information is available to this character at this point?

The full script is not the character’s context

Imagine a scene where Noor enters a station office to recover her brother’s bag. The audience saw a recorder sewn into its lining. Noor did not. A clerk knows the bag was moved upstairs, but has no reason to know about the recorder.

If a rewriting prompt includes the whole outline, it includes all three perspectives. A fluent output might let Noor say, “Give me the recording,” turning audience knowledge into character knowledge without a discovery.

That is a continuity error unless the writer intends to reveal that she knew all along. The problem is not the wording. It is the missing evidence for her knowledge.

Track fact, belief, and source separately

A small ledger can keep those perspectives apart. Here is an illustrative JSON record for the start of the scene:

{
  "scene": "station_office",
  "story_fact": "recorder_is_in_bag_lining",
  "audience_knows": true,
  "characters": {
    "Noor": {
      "status": "unknown",
      "source_scene": null
    },
    "clerk": {
      "status": "unknown",
      "source_scene": null
    }
  }
}

Use a separate record for the bag’s location. The clerk knows it moved upstairs. Noor believes it is still behind the desk. Her mistaken belief is not a data error to “correct”; it gives her a reason to ask the wrong question.

For each important fact, record only three things: who knows or believes it, what supports that state, and which scene changes it. Avoid cataloguing every incidental detail. Start with information that changes a decision, exposes a lie, or enables a reveal.

Make the review return evidence, not reassurance

Alongside the scene text, give a rewriting system a compact instruction:

Preserve the knowledge states supplied for the start of this scene. Do not let a character act on audience-only information. If a character learns something, identify the observable event that teaches it. After the rewrite, list any knowledge changes with a quote from the scene. Mark uncertain cases for review instead of inventing an earlier discovery.

This is a prompt specification, not a guarantee. A model may still miss a violation or produce an unsupported explanation. Review the cited line yourself.

A useful review output has a character, a fact, the before-and-after state, and the supporting line. “Continuity preserved” provides much less to inspect than “Noor now knows the bag is upstairs because the clerk says, ‘It went up with the other evidence.’”

Check the next scene as well

Suppose Noor opens the lining at the end of this scene. Her knowledge has changed, but finding a recorder does not mean she knows what it recorded. Those should be separate entries.

Now review the following scene. Does she search for a way to play it, or immediately accuse someone using information she has not heard? A local rewrite can create a downstream error even when its own ending works.

An intentional reveal is different. If Noor’s impossible knowledge is a clue, annotate that intention and its later explanation. Do not automatically reject every apparent inconsistency.

Start with one reveal

Pick a reveal in your current script. Work backward to the earliest moment each person could know it, then inspect the intervening dialogue. Keep the ledger beside whichever writing tool you use.

This is part of the broader story-workflow problem that interests me while building Laper: good prose still needs an accountable path from information to action. You can learn about the project at laper.ai, but the exercise needs no particular product.

A character can guess, bluff, or lie. The rewrite should give you enough evidence to tell which one is happening.