🧩大炮API Skill加载器@炮老师的小课堂

The word "Skill" gets thrown around a lot in AI-adjacent tools, but here it means something concrete: a packaged set of system-prompt logic and reference-loading rules that turns the dapaoAI chat workbench into a specialist. DapaoAPISkillLoaderNode is the node that scans, selects, installs, and manages those Skills - the entry point to the most distinctive thing this pack does.

What it does

The single required input, 🧩 Skill选择, is a dropdown that lists every installed Skill plus an 自动选择 (auto-select) option. Out of the box the pack ships 17: 3D动画短片生成器, Beautiful Article, 品牌宣传短片生成器, 电商详情页工作流 (detail-flow), GPT Image 2, H3视频提示词生成器, Music 3提示词重写器, a kb-retriever for local knowledge-base lookup, and more - each with a stable ID in brackets like [detail-flow]. It reads from the pack's own skills/ directory, and it's compatible with a sibling pack's ComfyUI-llama_Dapao/skills/ folder if you have that installed too.

Output is a single 🧩Skill配置 of custom type DAPAO_API_SKILL_CONFIG, which plugs into DapaoAPIMultiTurnChatNode's 🧩Skill配置 input.

The routing catch that decides your bill

Here's the part people miss. 自动选择 is not free - the tooltip is explicit that auto-selection makes one extra LLM call to the currently-configured API model for a deterministic routing decision. Pick a specific Skill by name instead and that call never happens. If you're building a repeatable pipeline, hard-code the Skill; reserve auto-select for exploratory work where you genuinely don't know which specialist fits.

There's a second, less obvious charge lurking: Skills that read references on demand can trigger "a clear two-stage call." So a Skill isn't just prompt text - it can pull in files and pay for a second request. Don't be surprised by two line items on your bill for one turn.

Installing your own Skills

The loader supports uploading a Skill folder or a ZIP: standard single-directory layout, or a full repo with 仓库/skills/*. The upload path has real hygiene - ZIPs are checked for path traversal, symlinks, duplicate paths, and oversized content, and same-name Skills don't auto-overwrite. That's noticeably more careful than the average "just drop it in" custom node, and in a category where a node that phones home with your key is exactly the attack surface people worry about (this KB's external-api-nodes.md keeps a sharp eye on that), it's a good sign.

You can also rename a Skill's display name without touching its content - aliases are labels only, and the stable ID is what old workflows keep referring to. The optional 🤖API模型 input exists only to power an "AI-optimize Skill display name" button; you can skip it and rename manually.

Workflow shape

Config (model/key) + Settings (how it talks) + SkillLoader (what it knows) + optional MaterialLibrary (what it can look at) all feed the chat node. A Skill effectively replaces your system prompt - the chat settings' own tooltip says when a Skill is connected, the Skill owns the system rules, so leave the settings system prompt empty.

Gotchas

• "所选 Skill 不存在" means the Skill list changed under the node - refresh the node, old enum entries go stale fast when you add/remove Skills.
• Auto-select burns a call every run. Budget-conscious users run manual.
• Two-stage reference loading can double a turn's latency - expect it, don't debug it as a hang.
• Chinese-only menus, as with the whole pack.

Install is standard: ComfyUI Manager → "dapaoAPI", or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git → pip install -r requirements.txt → restart. Skills live in ComfyUI-dapaoAPI/skills/ if you want to browse what shipped.