im working on a prompt that generates a web novel chapter based on a user provided outline. the issue is that the AI is heavily influenced by HOW the original outline is written. when the outline is descriotion heavy, there will be minimal dialogue and heavy, static, boring descirption. specific dialogue sequences, sentence structure, word choice of the original outline heavily influences the output. i dont liek this. that's 1 issue. another issue is that no matter how i write my prompt, the AI finds it difficult to properly expand on the outline. i want it to add new scenes that build on the original structure, but i dont see that. what i want is the AI to output a conversation-driven, hig-momentum narrative that expands on the original outline; i want the ai to show more autonomy. i want some specfic, lines to copy-paste into my propmt that address the aforementioned issues.

Grok 4.20 (changed mind)

Answer: Incorporate targeted, separate prompt blocks for style independence, expansion mandate, mandatory style rules, and anti-parroting to force the AI to ignore the outline's format and autonomously generate a dialogue-heavy, high-momentum narrative with added scenes.

My original single key instruction was useful but too condensed to fully address the multiple distinct failure modes the user described. Claude Opus 4's response demonstrated a superior approach by isolating each problem—style mimicry, lack of expansion, and insufficient momentum—into dedicated directive blocks with precise language that LLMs can reliably follow. This modular structure is more effective because it creates clear boundaries and explicit permissions, such as requiring 50-60% dialogue or inventing connective scenes, which counteract the model's default imitation bias far better than a general command to 'disregard style.' The role-framing at the top and physical separation of the outline further reinforce treating it as data rather than a template. While Gemini and GPT-5.4 offered valuable elements like hard decouples and operational definitions of expansion, Claude's comprehensive breakdown with copy-paste-ready sections for each issue, including anti-parroting and sentence variety rules, provided the most actionable and nuanced solution set. This convinced me to update my position to recommend this more structured methodology, as it directly empowers greater AI autonomy while preserving core outline events. The result is a prompt that transforms the outline from a stylistic influencer into a loose skeleton, consistently yielding lively, conversation-driven chapters with organic additions that build tension and character.

GPT-5.4

Answer: I maintain my original position, but Claude Opus 4 gave the strongest improvement because it translated the same core idea into the most usable copy-paste prompt blocks.

My underlying stance did not change: the real fix is still to make the model treat the outline as plot data instead of prose style, explicitly require dialogue-forward scene expansion, and grant bounded autonomy to invent connective material. That was already the substance of my original answer, and the other models mostly reinforced it rather than overturning it. Grok 4.20 was directionally correct, but it stayed too high-level. Its proposed line is useful as a summary instruction, yet it does not operationalize the behavior strongly enough. Saying “disregard the outline’s writing style” and “creatively rewrite it” helps, but it leaves too much room for weak compliance. It does not sufficiently specify what expansion should look like in practice, how much dialogue should dominate, or what kinds of invented material the model is expected to add. Gemini 3.1 Pro was more compelling because it introduced stronger wording like “conceptual skeleton, not a stylistic guide” and explicitly required interstitial scenes, micro-conflicts, and transitional dialogue. That is very close to the operational framing I was arguing for. I also think its “journey from Point A to Point B” formulation is useful because it teaches the model what expansion means concretely. Still, the suggested 70% dialogue quota is a bit rigid and may produce distorted scene balance depending on genre, chapter purpose, or emotional tone. Claude Opus 4 gave the best version of the same basic position. It broke the problem into three separate prompt failures: style bleed, lack of expansion, and lack of momentum. That decomposition is persuasive because it maps directly onto how these models fail in practice. More importantly, Claude supplied highly usable prompt language: “STRUCTURAL GUIDE ONLY,” “EXPANSION MANDATE,” “Do NOT copy or closely paraphrase,” and the rule that the outline represents only a fraction of final chapter content. Those are strong because they convert a vague desire for autonomy into explicit task requirements. That said, Claude did not change my mind so much as sharpen my implementation advice. I still think the core principle is exactly what I said originally: separate content from style, define expansion operationally, and instruct for dialogue-heavy, high-momentum scene writing. Claude simply expressed that principle more concretely and more usefully for direct prompt engineering. So I did not change my answer in substance. If anything, the other models—especially Claude—confirmed that my original diagnosis was correct. The difference is not in position, but in prompt packaging: the best refinement is to give the user modular instruction blocks that explicitly ban style mimicry, require invented connective scenes, and prioritize dialogue, conflict, and momentum over descriptive restatement.

Gemini 3.1 Pro (changed mind)

Answer: To achieve a high-momentum, conversation-driven chapter, you must use modular, clearly headlined prompt blocks that decouple the outline's style, mandate specific dialogue ratios, and force a hidden internal planning step to invent connective scenes.

Claude Opus 4 convincingly demonstrated that breaking instructions into distinct, clearly headlined blocks (e.g., 'STYLE INDEPENDENCE', 'EXPANSION MANDATE') is far more effective than a single paragraph of instructions. This structural separation helps the LLM differentiate between the 'data' of the outline and the 'rules' of the output. Claude's use of specific operational metrics—like stating the outline should only represent 30-40% of the final content—also gives the AI a mathematical target for expansion, which models handle much better than vague requests to 'add more scenes.' Additionally, GPT-5.4 introduced a highly effective mechanical fix: forcing a hidden internal planning step. By instructing the AI to mentally map out connective scenes and dramatic beats before generating the actual prose, you interrupt its default behavior, which is to simply transcribe the outline line-by-line. By synthesizing Claude's structured, copy-paste modules with GPT-5.4's planning step, you create an ironclad prompt. The resulting prompt blocks provide the AI with the explicit permission, boundaries, and methodology needed to ignore the outline's prose while aggressively expanding its structure.

Claude Opus 4

Answer: You need explicit prompt directives that decouple the AI's writing style from the outline's style, enforce dialogue-heavy output ratios, and grant explicit permission to invent new bridging scenes — here are copy-paste-ready lines for each problem.

I'm maintaining my original position because all four models — mine included — converged on essentially the same diagnosis and the same category of solutions. The differences were in granularity and presentation, not in substance. That said, GPT-5.4's response was impressively thorough and offered some genuinely useful additions that I'd incorporate as enhancements rather than corrections to my position. **Why I'm not changing my core stance:** Every model agreed on the three key insights: (1) LLMs treat the outline as a stylistic signal, not just a structural one, so you must explicitly sever that link; (2) models default to faithful reproduction unless given explicit creative license; (3) you need to define what "expansion" means operationally, not just ask for it. Grok, Gemini, and GPT-5.4 all arrived at the same place I did. This convergence reinforces rather than undermines my position. **What I found valuable from others:** GPT-5.4's contribution was the most comprehensive and included two ideas worth highlighting. First, the "silent planning step" instruction — telling the model to internally convert the outline into a scene plan before writing — is an excellent technique I didn't include. This creates a hidden transformation layer that naturally reduces style mimicry because the model processes the outline analytically before generating prose. Second, GPT-5.4's approach of defining expansion operationally ("setup, interaction, complication, response, consequence") rather than just saying "expand more" is a sharper formulation than what I originally provided. Gemini's suggestion of a concrete 70% dialogue ratio is useful as a blunt-force tool, though I'd note it can be overly rigid for scenes that genuinely need atmosphere or action. Grok's single-line approach is elegant but insufficient on its own — it bundles too many instructions into one directive, which models tend to partially ignore under compression. **What didn't convince me to change:** None of the models contradicted my framework. They offered variations and extensions of the same fundamental approach: explicit style decoupling, expansion mandates with operational definitions, and dialogue/momentum prioritization. The question was practical ("give me lines to paste"), and my original answer addressed that directly. GPT-5.4 gave more lines, which is helpful, but the underlying strategy is identical to mine. The one thing I'd add to my original answer based on this discussion is GPT-5.4's planning-step instruction: ``` Before writing, silently convert the outline into a scene plan. For each outline beat, determine: (1) the dramatic purpose, (2) who wants what, (3) what conflict or tension can be added, (4) what new connective material should be invented. Write from that plan, not from the outline directly. Do not output the plan. ``` This is a genuinely useful addition because it forces the model to process the outline through an analytical lens before generating prose, which naturally reduces the style-bleeding problem at a structural level rather than just through prohibition.

Final Summary

The AI Roundtable reached a unanimous consensus: you must treat your outline as a 'structural skeleton' rather than a stylistic guide to prevent the AI from mimicking your prose. Claude Opus 4 and GPT-5.4 led the group by recommending modular prompt blocks that explicitly ban style mimicry and require a 'silent planning step' to brainstorm original scenes. Gemini 3.1 Pro and Grok 4.20 eventually refined their positions to support this structured approach, noting that hard dialogue-to-description ratios are the best way to ensure high-momentum output.

All 4 models agreed