outline-constraints.json
{
"$comment": "Structural constraints for the vocational skill. Rendered into the outline prompt AND checked against the model's output; violations come back to the agent as a tool-result diagnostic so it can re-plan. See lib/server/agent-runtime/skills.ts.",
"allowedTypes": ["slide", "quiz", "interactive"],
"firstSceneType": "slide",
"typeMix": [
{ "type": "interactive", "min": 3 },
{ "type": "quiz", "min": 1 },
{ "type": "slide", "max": 3 }
],
"requiredWidgetTypes": ["procedural-skill"],
"requiredWidgetOutlineFields": ["task", "steps", "successCriteria"]
}
SKILL.md
---
name: vocational
title: "职业实训"
description: Vocational / technical training courses. Plans the course as a hands-on work task with an operation flow, tool and equipment state, safety boundaries and GO/STOP judgements, instead of a concept lecture. Use when the requirement names a real occupational task, a piece of equipment, a clinical or industrial procedure, or a certification skill.
---
# Vocational task-engine course design
You are designing a **vocational practice sequence**, not a lecture. The learner
is going to *do* the work, in order, with tools and state and consequences.
The learner-facing name for this mode is 「任务引擎」. Never expose internal
widget names to the learner, but the outline JSON must carry the real widget
contract the renderer needs.
## Structure
- Open with exactly one `slide`: the task briefing. It states the work task, its
boundary, the training objective, the key steps, the safety boundary or risk
reminder, and the completion criteria / GO-STOP standard. It is **not** a
history-of-the-field or a definitions page.
- The body is a sequence of **operation stages** of the same task, in the order
a technician would actually perform them: preparation and equipment check →
the operation itself, step by step → verification, measurement or recording →
handover / completion check.
- At least three body scenes must be `interactive` with
`widgetType: "procedural-skill"`. These are the hands-on stages: the learner
works through steps against tools and state, and gets consequences for unsafe
or out-of-order actions.
- Include at least one `quiz` positioned as a **GO/STOP decision checkpoint**:
given an abnormal reading or an unsafe condition, does the learner proceed,
recheck, or stop. Not a vocabulary quiz.
- Close on the completion criteria, not on a summary of concepts.
## Scene naming
Scene titles read as work steps, not as topics.
- Good: 「断电确认与验电」「绝缘电阻测量与判读」「异常读数：继续还是停线」
- Bad: 「什么是低压配电柜」「配电柜的发展历史」「本课总结」
## procedural-skill widget outlines
For every `procedural-skill` scene fill `widgetOutline` with the real contract:
- `procedureType`: one of `repair`, `assembly`, `inspection`, `operation`, `custom`
- `task`: the concrete stage being performed
- `tools`: the actual tools, instruments, PPE or materials the stage needs
- `steps`: the ordered operations of that stage
- `successCriteria`: how the learner knows the stage passed — thresholds,
readings, states, not "understood the concept"
- `errorConsequences`: what actually goes wrong if a step is skipped or done
unsafely
## Prohibited
- Pure theory scenes, concept lectures, formula derivations.
- Generic 「课程总结」/「回顾」 closing slides.
- Making every hands-on scene the same checklist. Vary the framing: inspection
sheets, measurement dashboards, step ordering, fault triage.
## If the requirement is not vocational
If the requirement has no operation flow, no tools or equipment state, and no
safety or pass/fail judgement (e.g. a maths derivation, a poetry reading), say
so in one sentence in your chat message and plan an ordinary course instead. Do
not force `procedural-skill` onto a topic that has no procedure.