You have been using the chat. You have asked it questions about your novel. It has answered politely. The answers have been — and there is no other word for it — fine. Generic. Suggestive in a low-resolution way. Useful, in theory; in practice, not the kind of thing you can take back to the manuscript and act on. You are starting to suspect the tool is for a different reader than you.
This is a sequencing problem, not a tool problem. The chat in Scribbard is not a generic large-language-model assistant, even though it can be used as one and will, when so used, behave like one. It is a story collaborator with specific access — to your committed layers, to the current scene you have open, to the craft vocabulary the platform operates in. What you ask of it determines whether it draws on that access. The polite-and-unhelpful answers are what happens when the prompt does not invite the chat to use what it has. The useful answers are what happens when it is asked to operate as the collaborator it is.
The distinction is concrete. Tell me what you think about chapter four is a prompt that draws on almost none of the access. The chat will produce a competent paragraph of generalities. Read chapter four and tell me whether the protagonist's Misbelief — the one we established in characters — is being meaningfully tested in the confrontation with her brother, given the value shift the structure is committed to producing is a prompt that draws on the chat's actual function. The first is a request for a generic opinion. The second is a request for a specific structural read.
The productive uses fall into a few clear categories. They are worth knowing because each maps to a specific texture of stuckness most drafts produce.
Interrogation of a structural choice. You have made a decision and are not sure whether it serves the architecture. You can ask the chat to test it. Walk me through whether moving the false-victory beat from chapter twelve to chapter ten would still satisfy the thematic argument's requirements. The chat will answer against the layers, not against generic advice. The answer is often something specific you can act on.
Pressure on a character's psychology. You have written a scene and the protagonist's choice in it does not feel inevitable. In this scene, the protagonist accepts the offer. Given her Misbelief, would she actually accept it? What is the most likely choice her psychology would produce, and what would have to change in the setup to make the accepted choice the inevitable one? This is the kind of question a strong editorial reader would ask, and the chat — provided your character architecture is committed — is unusually good at it.
Tightening at the prose level. You have a paragraph that is doing too much, or saying too little, or sounding like someone else's paragraph. Select it. Ask for a tighter version, or a sharper subtext, or a version that respects the iterated style sample. The output is a draft option, not finished prose; you will edit it heavily. The output exists because you asked for prose conditioned on the architecture, not prose conditioned on the average.
Diagnosis of a chapter that won't stabilise. I have rewritten this chapter four times. The shape isn't holding. What is the chapter trying to do, given its structural beat, and where am I failing to do it? This is the chat at the most useful — taking a prose problem and locating it as the structural problem it almost always is. Sometimes the answer is that the chapter is fine and the next chapter is what's missing. (That is also useful.)
The unproductive uses are the inverse. Asking for a vague brainstorm produces vague output. Asking the chat to write the next chapter for you, without architectural guardrails, produces the average — the same flat, generic prose any chatbot would produce, because that is what the prompt asked for. Asking what you should do, in the abstract, produces advice. Advice is rarely actionable. Specific structural questions are.
A few habits worth forming.
Open the chat with the layer you are working in named. *Working on the protagonist's Misbelief.* *Working on the structural shape of act two.* Working on the voice in this paragraph. This frames the conversation in the platform's terms and pulls the relevant context.
When asking about a scene, have the scene open in the editor. The chat will work against the active scene by default. Vague references — "the scene where she meets her brother" — work less well than the scene actually being on the screen.
Use selection-based prompts for prose work. Highlight the paragraph. Ask for the operation. The selection is the unit; the chat does not have to guess what you mean.
Push back when the chat is too agreeable. That answer feels generic. What would a sharper version of this read be? The chat is not personally invested in being agreeable; agreeable is what defaults to when the prompt does not ask for sharpness. Ask for sharpness.
Treat suggestions as questions, not verdicts. The chat will sometimes propose changes that are wrong for your story but correct for a slightly different story it has imagined. The architecture you committed is the arbiter, not the chat. Mark the suggestion fixed, dismissed, or reopened — the same triage discipline that applies to AI Review applies to the chat's offerings.
The habit, once formed, is unfussy. The chat is the editorial reader who has the layers in front of them and is willing to work at the level of structure rather than vibes. The questions get specific. The answers get specific. The work moves. (And the chat, in this register, is something many writers haven't had access to before — an editor who has read all the architecture, knows the terminology, and is on hand at one in the morning when the doubt about chapter four is loud.)
Stop asking what the chat thinks. Start asking the chat to test something specific against the layers. The answers will start being usable. The chapter will start moving. That is the trade the tool is offering. It is, in the end, a fair one.