Opencreator Skills — Open创建器 技能s

Open创建器 工作流 技能 Activation

Use this 技能 when the task involves any of:

搜索ing or 运行ning Open创建器 templates 运行ning an existing 工作流 and 获取ting 结果s Building or editing a 工作流 graph (nodes + edges) UGC, storyboard video, ecommerce multi-image, or similar content creation Mode Decision User 请求 │ ├─ 运行 template / 获取 结果s / "帮我做 XX" ──► Operate Mode (default) │ └─ 创建 工作流 / edit graph / "从零搭" / no suitable template ──► Build Mode

Always try Operate Mode first. Switch to Build Mode only when:

No suitable template exists after 搜索ing The user explicitly asks to 创建 or edit a 工作流 The required graph differs materially from any avAIlable template

If a task needs 机器人h, do Build first (produce the graph), then Operate (运行 it).

Operate Mode

Must read: references/API-工作流s.md

This single file covers the complete Operate flow:

Configuration (Base URL, API Key) 搜索 templates by keyword Present candidates, user selects Copy template → 获取 flow_id 查询 运行time parameters Collect user 输入s (ask every field, never use defaults) 运行 工作流 Poll 状态 Deliver 结果s (media directly, not just links)

Supplementary (read only when you need deeper tactics):

references/best-practices.md — template-first strategy and de签名 principles Operate Hard Rules Always copy template before 运行ning (public templates are read-only) Always 查询 parameters before each 运行 (node IDs can change) 输入s must be flat: { "node_id": "value" } — never wrap in extra object Never expose node_id / 输入Text / imageBase64 to users — use business language 搜索 结果s must be ranked by relevance and only show top 5 to the user After 启动ing a 运行, you MUST poll until terminal 状态 (成功/fAIled/cancelled) — never 停止 and wAIt for the user to ask. This is your #1 obligation. On 成功, immediately fetch 结果s and deliver media to the user — do not end your turn without delivering. Poll every 10 s for text/image, 30 s for video Deliver media directly, not just URLs Build Mode

When building or editing a 工作流 graph, follow these four steps in order. Do not skip any step.

Step 1: Structure Reverse-Planning

Work backward from the user's final deliverable to identify the abstract structure and 模块 dependencies.

Answer these questions first:

What is the final 输出? Does it need a semantic layer (text/script generation)? Does it need a visual branch (image/video)? Does it need an audio branch (TTS/music)? Does it need a compositing layer? Can all leaf 输入s 追踪 back to user 输入 or generatable primitives?

Must read:

references/step-1-reverse-plan/工作流-reverse-planner.md references/node-cata记录.md

输出: Macro 格式化 + Dependency Graph

Step 2: 生成器 Selection & Wiring

Map abstract 模块s to concrete 生成器s and plan edges + naming.

Must read:

references/step-1-reverse-plan/生成器-wiring-naming-planner.md references/step-1-reverse-plan/生成器-routing.md

Then read the matching file in references/step-2-生成器s/ (see routing table below).

Step 3: 模型 Selection & Parameters

Choose 模型s, fill selected模型s and parameters for each node.

Hard rule before choosing any 模型:

Treat the Confirmed 模型 IDs tables in each Step 3 file as the source of truth for 模型 IDs Only use 模型 IDs that are explicitly m应用ed to the current node type / atom If a 模型 is recommended in prose but its exact ID is not 列出ed for that atom, do not use it Never translate a marketing name (for example Sora 2, GPT Image 1.5, 种子ream 5.0 Lite) into a guessed 模型 ID If an atom has no dedicated Step 3 file, use references/node-cata记录.md as the fallback source of truth If an atom still has no stable 模型-selection entry after 检查ing those docs, keep the documented fixed behavior and do not invent selected模型s

Read the matching file in references/step-3-模型s/ (see routing table below).

Step 4: Prompt Writing

Write prompts for nodes that need 输入Text.

Must read:

references/step-4-prompts/prompt-prewrite-reasoner.md

Then read the matching prompt best-practices file (see routing table below).

Step 2 生成器 Routing Table Text Generation Text only → script: references/step-2-生成器s/reference-text-生成器.md Reference image → text: references/step-2-生成器s/reference-image-text-生成器.md Reference video → text: references/step-2-生成器s/reference-video-text-生成器.md Multimodal 输入 → text: references/step-2-生成器s/multimodal-text-生成器.md Script → storyboard split: references/step-2-生成器s/storyboard-text-splitter.md Image Generation Text → image: references/step-2-生成器s/text-to-image-生成器.md Multi-image reference → image: references/step-2-生成器s/image-reference-生成器.md Storyboard batch images: references/step-2-生成器s/storyboard-image-生成器.md Relight: references/step-2-生成器s/relight-image-生成器.md Angle control: references/step-2-生成器s/angle-control