✍️04漫剧剧本创作@炮老师的小课堂

Every AI drama pipeline eventually has to write the actual script, and this is where the 🐠 漫剧 suite does it. DapaoDramaWrite (advanced node 04) produces the single-episode 剧本.md - the doc that every visual stage downstream (assets, storyboard, image prompts, video prompts) is built from. Get this right and the whole chain is happy; get it sloppy and you'll be paying to regenerate everything after it.

What it actually does

Three jobs, and the third is the one that distinguishes it from a generic "write me a script" LLM prompt:

1. Draft from scratch - one episode at a time, driven by the series plan from node 03 and your "📝 本阶段要求".
2. Normalize an existing script - paste in a finished episode you already have (maybe written by a human, maybe generated elsewhere) and it reformats and tightens it while preserving plot, characters and key dialogue. Its instruction for this mode is explicit: normalize, keep the story, flag contradictions - don't rewrite the story's direction on your own.
3. Targeted revision - hand back a script through "📄 原文或现有稿" with specific notes, and it edits only what you asked.

The output format is a hard contract, not a suggestion: scenes are headed ## EP001-SC001 style, with 场次 (scenes), 动作 (action), 对白 (dialogue), 声音 (sound) and the episode's 交接 (hand-off state) preserved. Every downstream mechanical check keys off that ## EPxxx-SCxxx heading format, so a write stage that ignores it poisons everything behind it.

How it works

It's one of the eight LLM stages, so the plumbing is shared: calls go to the author's relay (api.dapaoai.com/v1/chat/completions), the 🔑 API密钥 field or DAPAO_API_KEY env var authenticates you, and it runs under a long system contract that forbids the model from doing anything but this stage. The 📄 系列开发.md input from node 03 feeds it the plan. On the way out, the node checks the structure (right scene headings present, no promised-but-missing pieces) and returns the 📦 漫剧资料包, the 📄 阶段文档, the ℹ️ 处理信息 report, and an empty 🧩 提示词列表 (this stage writes prompts for nothing - that's nodes 06–08's job).

Inputs that matter

• 🤖 LLM模型 - pick from the unified model list (defaults to gemini-3.7-flash). Note the model dropdown is shared across the whole pack, so you can put a big-arc planner here and a cheaper model elsewhere.
• 📝 本阶段要求 - where you steer: tone, characters to feature, what must happen this episode, what to seed for later.
• 📄 原文或现有稿 - paste source or an earlier draft for normalize/revision mode.
• 🎲 随机种 - fixed by default, which is the key to the 48-hour request-reuse cache: same content + same seed = not re-billed on re-run.

Installing and gotchas

Pack install applies: Manager → "dapaoAPI", or clone the repo into custom_nodes and restart. No model downloads; the spend is per-LLM-call on the dapaoAI relay, and failed paid requests don't auto-retry.

The trap to avoid: don't hand it the whole novel in one go and expect a full season. The suite is strictly one episode at a time - the novel stage segments the source for analysis, but the write stage drafts the current episode, and long multi-episode source should be given per-episode. If you paste a finished human script, keep the "已有剧本" workflow in mind - the guided DapaoDramaPrepare node handles the normalize path (plus optional novel analysis and series planning) in one go, so you only drop to this advanced node when you want the write stage isolated for fine-grained revision.