vibecanvas-image-generate

name	vibecanvas-image-generate
description	Generate fresh images on the VibeCanvas canvas using the OpenAI-compatible Image 2 API. Use when the user asks to create, produce, draft, imagine, or render a brand-new image (not edit an existing one). Covers single-shot generation, multi-candidate generation, and Prompt Architect-driven brief expansion.
license	MIT
compatibility	opencode, zcode, mcp
metadata	{"product":"vibecanvas","role":"image-generation","protocol":"2.0"}

VibeCanvas Image Generate

Use this skill when the user wants to create a new image from scratch — text-to-image, multi-candidate generation, or expanding a short brief into a polished prompt. For modifying an existing image (source/reference/mask/annotation), route to vibecanvas-image-edit.

Preconditions

1. The VibeCanvas service is registered as MCP. Verify with get_workspace_context {}. If it errors, ask the user to run install-skills and restart the session.
2. An image provider is configured. Call get_provider_capabilities {}. Read configured — if false, the provider token is missing and generation will fail at run time. Tell the user to set IMAGE_API_KEY (or edit the provider profile via the Web UI's Provider panel) before continuing.
3. Read the candidate limits. From the same capabilities response, note maxCandidates and the capabilities object (e.g. batchN, customSize). Don't request a candidateCount higher than maxCandidates.

Workflow

1. Inspect the current selection.

   • get_selection_context {}.
   • If the user pointed at an existing input.brief node, read its config.text — that's the creative brief.
   • If the user pointed at an agent.prompt-architect node, they've already authored the prompt spec; skip to step 4.

2. Ensure a brief exists.

   • If no brief is selected and the user gave you text, create one with a patch:

     { "op": "addNode", "node": { "id": "node-brief-01", "type": "workflow", "position": { "x": 80, "y": 200 }, "data": { "nodeType": "input.brief", "config": { "text": "<the brief>" }, "status":    
     

Prompt design (when authoring PromptSpec directly)

If you're feeding image.generate.prompt without going through agent.prompt-architect (e.g. via updateNode with a hand-written spec), use this shape:

{
  "subject": "Concise subject line (<= 160 chars)",
  "purpose": "What the image is for",
  "composition": "Framing, perspective, hierarchy",
  "lighting": "Direction, quality, color temperature",
  "materials": ["..."],
  "palette": ["..."],
  "style": " Photorealistic / editorial / illustration / ...",
  "avoid": ["watermarks", "editor chrome", "annotation arrows", "garbled text"],
  "finalPrompt": "The actual prompt sent to the image model — combine everything above into one coherent prompt."
}

Only subject and finalPrompt are required; other fields guide the LLM when agent.prompt-architect runs.

Notes

• image.generate writes the generated images as candidate Artifacts under .vibecanvas/artifacts/. Their parentArtifactIds point at any reference images used.
• Provider billing accrues per request, even on cancellation. If the user cancels mid-generation, cancel_run { runId } propagates an AbortController to the in-flight HTTP request, but the provider may still charge for compute already started.
• For aspect ratios, prefer the utility.aspect-ratio node over hard-coding { width, height } in image.generate.config — the utility node emits a typed AspectRatio value that the runner normalizes to the provider's supported sizes.
• If the user wants the same prompt at multiple sizes, build parallel image.generate nodes fed by one agent.prompt-architect; don't try to fan out inside a single node.

See references/provider-contract.md for provider profile details.