🛍️电商详情页提示词@炮老师的小课堂

E-commerce detail pages are a weird niche with a real workflow: a product page is a sequence of screens - hero, selling points, specs, social proof, CTA - and the images need to stay visually continuous across all of them. DapaoLocalDetailFlowPromptNode ("电商详情页提示词") generates that whole sequence as prompts in one shot: 1–8 screens, from first-frame opening to a closing CTA, with continuity enforced by the prompt contract rather than left to chance.

This is another of the seven dapaoAPI tools ported to local GGUF inference. The original was an API node; the local version keeps the entire input/output layout and redirects the OpenAI-compatible request to your own llama.cpp model. No API key, no cloud - and unlike the API original, nothing is filtered.

The inputs that matter

• 🧭 页面类型 - 实体商品详情页, AI工具/模型产品页, SaaS功能页, 开发者产品页, or 自定义长页. Sets the genre.
• 🗂️ 产品品类 - auto-detect plus 食品饮料, 美妆个护, 3C数码, 家居家电, 服装箱包, 母婴用品, 宠物用品, 汽车用品, 运动户外, 珠宝饰品, AI软件/模型…
• 🛒 适配平台 - 淘宝/天猫, 京东, 抖音小店, 拼多多, 小红书, 独立站, or 通用长图. Different platforms have different image conventions.
• 🔢 分屏数量 - 1 to 8. The tooltip is the design philosophy: no matter the count, it always runs from opening screen to ending CTA - a 1-screen page still gets a complete arc.
• 🔗 页面连续方式 - 严格连续详情页 (recommended) vs 统一视觉、分屏独立构图.
• 📝 原始电商需求 - your product brief. Everything else is scaffolding.
• Then the optional refinement bank: 🛍️ 产品名称, 📝 真实参数与产品事实, 👥 目标购买人群, 💡 主卖点方向 (with second/third selling points), 📏 主要证据类型, 🎬 主要使用场景, 🛒 CTA类型, 🎨 视觉风格方向, and reference images - 📦 产品图1 through 📦 产品图6 (identity facts) and 🎨 风格参考图1–🎨 风格参考图3 (style DNA only).

The outputs

A wall of them, and each has a job: 🧭 可选视觉母版提示词 (the long-poster "mama board" master prompt), 🖼️ 第01屏提示词 through 🖼️ 第08屏提示词 (per-screen), plus three connection formats the description advertises - 📚 所选分屏多行提示词, 🧩 所选分屏批量提示词 (a STRING list, for batch ComfyUI workflows), 📋 详情页完整蓝图JSON, 📝 全部准确画面文案, ⚠️ 事实与连续性提醒, 📄 LLM完整响应, and ℹ️ 处理信息.

Where people get burned

This node's max output is 12000 tokens by default and requests can legitimately take a while on a small local model - if it times out, raise ⌛ 请求超时 (default 600s) before blaming the node. Also: 🔄 每次重新生成提示词 defaults to off, so the node reuses ComfyUI's cache and you'll get the same prompts on re-run; flip it on when you actually want a fresh roll. And remember 🎲 随机种 only controls ComfyUI caching, it isn't sent to the LLM.

Installing it

Standard pack:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-llama_Dapao

then requirements + restart. Pack-wide gotchas: llama-cpp-python from the JamePeng fork at 0.3.47+ (stock PyPI 0.3.35 errors on Qwen3.8 GGUFs), NVIDIA GPU wheel overlay, models in ComfyUI/models/LLM/. For the reference-image analysis you need a vision model + mmproj; pure text briefs work with any GGUF.