AI Writing Context: Chat vs. Scene Generation and How to Troubleshoot Prompts
Learn what context an AI writing assistant needs, when to use chat or scene generation, and how to troubleshoot vague, contradictory, or overlong fiction prompts.
Use chat to explore decisions and scene generation to produce a bounded passage. Both workflows need the same core context—viewpoint, current story state, goal, resistance, and limits—but they should return different kinds of output.
What context does an AI writing assistant need?
Give the assistant the smallest packet that can change the answer:
| Context | Example |
|---|---|
| Current state | Mara has the letter but not the key |
| Goal | Get access to the storeroom tonight |
| Viewpoint | Close third person through Mara |
| Relevant facts | The director controls the archive and suspects her |
| Boundaries | Do not reveal the mother's fate |
| Desired next state | Mara gets the key but creates a new risk |
Separate established facts, current uncertainties, and possible ideas. This prevents a suggestion from being mistaken for canon.
When should I use chat?
Use chat for questions that still have multiple reasonable answers:
Given the premise and scene state below, propose three ways the director could resist.
For each option, explain the pressure it creates, what Mara must choose, and the next scene problem.
Do not write prose and do not select an option for me.
Chat is good for alternatives, diagnosis, research questions, and continuity review. Its output is usually a set of decisions to evaluate, not manuscript text to paste.
When should I use scene generation?
Use scene generation after you know what the passage must accomplish:
Write approximately 400 words of the key exchange.
Use close third person through Mara and past tense.
She wants the storeroom key; the director wants the original letter.
Preserve: she has no proof of his crime.
End: he gives her the key, but notices the envelope is too light.
Output prose only. Add no characters or backstory.
Scene generation is not a substitute for deciding the scene's purpose. If the output wanders, check the goal and ending before adding more style adjectives.
Troubleshoot the four common failures
The output is generic
Replace “make it emotional” with an observable constraint: “show anger through an interrupted action and one avoided question.” Name the character's objective and what they refuse to say.
The output contradicts the manuscript
List the conflicting facts explicitly and tell the assistant which one is authoritative. Then reduce the context to the current scene and check whether an old note is being treated as current.
The output reveals too much
Add a knowledge boundary: “The viewpoint character suspects X but does not know Y. Do not confirm Y.” Set the ending before the reveal you want to protect.
The output gets lost in background notes
Remove notes that cannot affect the scene. Keep a short project reference for stable facts, then pass the current scene state and relevant entries for the request.
What can go wrong?
Is more context always better?
No. More context can introduce irrelevant facts, contradictions, and competing instructions. Context should be sufficient for the decision, not a dump of the entire manuscript.
Should I ask chat to rewrite a scene?
You can, but state whether the task is diagnosis, alternatives, or prose. For a rewrite, define what must remain unchanged and what may change; otherwise an improved sentence may also change story facts.
How do I know whether a prompt worked?
Check the output against observable criteria: correct viewpoint, preserved facts, intended turn, and ending state. “Sounds good” is useful, but it is not a continuity test.
Turn this guide into your next scene
Keep your outline, character notes, and draft together in NovelKnow. Start with one scene and review each AI suggestion before keeping it.