Claude Code Plugins

Community-maintained marketplace

Feedback

Create or refine a VibeDev Studio UnifiedWorkflow graph via tool calls. Use when the user says `$vd-create`, asks the model to “build the workflow graph”, “set up Research/Planning/Execution/Review steps”, or wants the model to collaboratively design prompts/conditions/breakpoints and write them into `unified_workflows` so the UI canvas shows the plan before running `$vd-next`.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name vd-create
description Create or refine a VibeDev Studio UnifiedWorkflow graph via tool calls. Use when the user says `$vd-create`, asks the model to “build the workflow graph”, “set up Research/Planning/Execution/Review steps”, or wants the model to collaboratively design prompts/conditions/breakpoints and write them into `unified_workflows` so the UI canvas shows the plan before running `$vd-next`.

vd-create (Workflow Authoring Assistant)

Goal

Turn a user’s project idea into a visual, editable workflow graph in the Studio:

  • four phases: research, planning, execution, review
  • node types: PROMPT, CONDITION (hard/soft), BREAKPOINT
  • explicit edges (nextStep, and for conditions onPass/onFail)
  • each prompt step includes: role/context/task/guardrails/deliverables + logInstruction

This skill is for authoring a UnifiedWorkflow graph — not for executing steps.

Key tools

Use the unified workflow tools (these update job_ui_state.graph_state_json.unified_workflows):

  • workflow_unified_get
  • workflow_unified_set_flow_fields
  • workflow_unified_upsert_step
  • workflow_unified_connect
  • workflow_unified_delete_step (rare)

Authoring workflow (no fluff, concrete)

1) Intake (ask only what you need)

Collect:

  • What are we building? (1–2 sentences)
  • User technical level: non-technical | some coding | experienced
  • Constraints: language/framework, OS, timeline, “must-have” vs “nice-to-have”
  • Definition of “done”: what can the user do when it’s complete?

2) Decide a graph shape (recommended default)

Create 3–6 steps per phase. Keep each step one prompt.

Recommended minimal skeleton:

  • Research: scope + repo scan (findings), risk scan (mistake ledger), breakpoint (memory pack)
  • Planning: architecture choices, step plan draft, breakpoint (handoff memory)
  • Execution: implement incrementally, add tests, condition (tests pass), breakpoint (handoff)
  • Review: run full verification, condition (all gates), final summary

3) Write the graph into the job

For each phase:

  1. Call workflow_unified_set_flow_fields to set name, description, global_context.
  2. For each node, call workflow_unified_upsert_step with a full step object.
  3. Call workflow_unified_connect to wire nextStep / onPass / onFail.

Step object requirements

Match the Studio schema:

  • id: stable string (recommend step_<phase>_<slug>)
  • type: PROMPT | CONDITION | BREAKPOINT
  • label: Step1, Step2, Sub1, etc.
  • title: human readable
  • colIndex: integer
  • subThreadLevel: integer
  • isSubThread: boolean

PROMPT fields:

  • role, context, task, guardrails, deliverables, logInstruction

CONDITION fields:

  • conditionType: script (hard) or llm (soft)
  • conditionCode: for llm, a TRUE/FALSE question; for script, a command line to run
  • onPass, onFail (or use workflow_unified_connect with edge=on_pass/on_fail)

BREAKPOINT fields:

  • reason
  • carryForward (memory prompt — instruct model to save a memory pack to context, not print it)

4) Confirm and hand off to the user

  • Tell the user which job/phase you updated and that it’s visible in the Studio.
  • Encourage them to edit anything, then start execution with $vd-next.

Guardrails

  • Keep steps small and explicit.
  • Prefer conditions as separate nodes (don’t embed “gates sections” in prompts).
  • Breakpoints should instruct “save memory pack to context” and “start a new thread”.