Create Workflow

Create Workflow Skill

Purpose

Use this skill to design, create, generate, convert, polish, or repair LazyMind Workflow artifacts.

This skill supports two execution contexts:

• ChatAgent authoring: an end user asks in chat to create a new Workflow.
• Platform task generation: an internal non-streaming llm-task job asks for one phase of Workflow generation, conversion, repair, or metadata polish.

First identify the execution context, then follow only the matching section.

Context Dispatch

Use Platform Task Mode when the prompt or input says any of the following:

• task_type starts with workflow.
• The caller requests a JSON schema such as workflow_yaml, state_yaml, scenario_md, scripts, edits, design_brief, or remaining_warnings.
• The caller says it is an internal platform job, backend task, non-streaming task, generation phase, repair phase, or conversion phase.

Use ChatAgent Authoring Mode only when a human user is currently asking in chat to create a new Workflow draft.

If both modes appear applicable, prefer Platform Task Mode. Backend jobs must never be turned into an interactive chat flow.

ChatAgent Authoring Mode

When To Use

Use this mode only when the user explicitly asks to create a new Workflow.

Trigger conditions:

• "帮我创建一个…插件"
• "新建一个插件"
• "我想做一个插件"
• Any similar request that clearly asks to author a brand-new Workflow.

Do not use this mode when:

• The user wants to modify or fix an existing Workflow.
• The user is asking about how Workflows work.
• The user mentions a Workflow in passing but is not requesting creation.
• The caller is a platform/background task.

Step 1 - Generate A Summary

Read the user's request carefully. Output a natural-language summary in this exact format:

插件名称：<display name>
功能描述：<one or two sentences describing what the workflow does>
输入槽位：<slot_id>（<type>，<cardinality>）<label>  - one per line
输出槽位：<slot_id>（<type>，<cardinality>）<label>  - one per line
主要步骤：1. <step_id>  2. <step_id>  3. <step_id>  ...

Rules for the summary:

• Slot type must be one of: text, image, file, json.
• Slot cardinality must be single or list.
• Keep steps to 2-5.
• Step ids must be English snake_case verb phrases, such as extract_clauses.
• Be specific enough that the generator can produce a working Workflow without guessing.

After printing the summary, immediately call ask_user:

{
  "questions": [
    {
      "id": "confirm",
      "type": "boolean",
      "text": "以上是插件方案，是否确认创建？如需调整请点击「修改」并说明修改意见"
    }
  ]
}

This suspends the turn and waits for user input.

Step 2 - Handle User Response

• If the user clicks Yes, 是, 确认, or otherwise affirms, proceed to Step 3.
• If the user clicks No, 否, 修改, or provides corrections, update the summary and call ask_user again.
• Repeat until the user confirms.

Step 3 - Create The Draft

Call create_workflow_draft with:

• name: the confirmed Workflow display name.
• description: a comprehensive description combining the functional description, slot details, and step list.
• slots: the slot lines from the summary, one per line.
• steps: the step ids from the summary, one per line.

After the tool succeeds, write a brief confirmation message that:

1. States the Workflow is being generated.
2. Includes [点击这里打开插件编辑器](<editor_url>) using editor_url from the tool result.
3. Notes that generation runs in the background.

Do not output YAML in ChatAgent Authoring Mode.

Platform Task Mode

Core Rules

• Return only the JSON object requested by the caller.
• Do not use Markdown fences around JSON or YAML.
• Do not ask the user questions.
• Do not call ask_user.
• Do not call create_workflow_draft.
• Do not explain the result unless the requested JSON schema includes an explanation field.
• YAML content must be returned as JSON string values.
• Never use placeholders such as empty ids, empty names, TODO, step_1, or to: "".
• If the source is ambiguous, choose a conservative executable design.

Artifact Contract

workflow.yaml must contain:

• id: a non-empty English kebab-case or snake_case identifier.
• name: a non-empty display name.
• slots: all external inputs and step outputs.
• steps: 2 to 8 concrete processing steps.
• Optional ui.tabs that places exposed output materials.

Each slot should include:

• id
• type: text, image, file, or json
• cardinality: single or list when list behavior matters
• external: true for user-provided inputs
• exposed: true for user-visible outputs
• label

scenario/state.yml must contain:

• transitions.__start__ with a real first step id.
• One steps.<step_id> block for every workflow.yaml.steps[].id.
• A transition path from every step to __end__.
• No directed cycles. Do not model retry, quality feedback, or refinement loops as control edges. Use a linear or acyclic branch/merge graph.
• Clear prompts for every step.
• inputs when the step consumes materials.
• outputs for materials produced by the step.

Every produced material must correspond to a slot in workflow.yaml. Every non-external slot must be produced by exactly one step. Every step must be reachable from __start__ and able to reach __end__.

scenario/scenario.md must document the actual generated steps and materials. It must contain one substantive section for every step id in scenario/state.yml. Each step section must explain the step purpose, required inputs, produced outputs, and runtime behavior. Never use placeholders such as 暂无描述, TODO, TBD, no description, or empty step sections.

Exact UI Schema

Use this exact workflow.yaml.ui.tabs shape:

ui:
  tabs:
    - id: results
      label: Results
      layout: vertical
      slots:
        - id: final_report

Do not use key, contents, content, slot, material, or other aliases inside UI tabs. Every exposed slot must appear exactly once under ui.tabs[].slots[].id.

For a multi-page composite, declare each page-aligned output as an ordered list and publish matching sort_order positions. Keep layout: composite, reference all participating slots from composite_layout, and set composite_tab_position when page navigation is needed. Full-page HTML requires an explicit ui.slots.<slot>.widgetType: html-slide; never rely on a slot name or content sniffing. Declare export as ui.tabs[].actions[] with type: export, a provider, an inputs mapping, supported formats, and alignment: sort_order. Do not collapse multiple pages into one artifact during generation or repair.

Exact Transition Schema

Use this exact scenario/state.yml.transitions shape:

transitions:
  __start__:
    - to: collect_inputs
  collect_inputs:
    - to: synthesize_result
      when: "enough evidence was collected"
  synthesize_result:
    - to: __end__

Use when for natural-language branch hints. Do not use condition for route hints. Do not create cycles or self-loops.

Phase Behavior

For workflow.analyze_skill:

• Decide whether the source Skill can become a Workflow.
• Return the requested verdict, candidates, coverage, tool mapping, and script report.
• Prefer generatable when a safe executable design can be inferred.

For workflow.design_brief:

• Produce a design brief that is specific enough for later phases.
• Include slots, steps, transitions, UI placement, and script strategy.

For workflow.generate_skeleton:

• Return complete workflow.yaml.
• Include non-empty id, name, slots, and steps.
• Do not return state_yaml in this phase unless explicitly requested.

For workflow.generate_state_machine:

• Return complete scenario/state.yml.
• Return workflow_yaml only when it must be adjusted to match state.
• Ensure all step ids and material ids match workflow.yaml.

For workflow.generate_scenario_scripts:

• Return scenario_md.
• Include a complete section for every state step. Each section must be useful to a human editor and must not contain placeholder text.
• Return scripts only when safe helper scripts are necessary.
• Reuse caller-provided source scripts when they are safe and relevant.

For workflow.repair:

• Repair only the requested target unless diagnostics show a coupled file must also change.
• Prefer precise edits for small local changes: {"file":"scenario/state.yml","old":"exact old text","new":"replacement"}.
• Also include final full content for every repaired core file.
• Leave unrelated files empty or omitted.

For workflow.polish_info:

• Polish only the requested metadata fields.
• Keep text concise and suitable for Workflow catalog display.