Image Generation

Image Generation

Use image_generation for image creation and editing. Keep every call in the typed envelope { "request": { ... }, "reason": "..." }.

Workflow

1. Choose generate when the result is based only on text. Send request.operation="generate" and a complete request.prompt. Set request.sizePreset only when the user requests a supported output size.
2. Choose edit when an existing image must influence the result. Send request.operation="edit", a complete request.prompt, and exactly one authorized request.inputPath.
3. Copy the exact image path returned by the producing tool or supplied by the user into request.inputPath. This includes workspace paths, authorized absolute/system paths, attachment readPath values, generated image-artifact://... paths, and revision-bound skill://... paths. For an attachment, call attachments_list first and copy its exact readPath. Never build a source object, guess an attachment path, or substitute a display name, ID, or private saved path.
4. Write reason as a short, non-empty, user-readable sentence describing the purpose of that specific generation call. It is display and audit metadata only and never grants access.
5. Inspect the returned status and Artifact contract. Report success only when status is succeeded and the returned Artifact is verified.
6. On success, the model result returns one top-level path, normally an application-owned image-artifact://... reference. Copy that same path into read_image.path to inspect it, or into a later image_generation edit request's inputPath to edit it again. Do not build a source object, combine URI and filesystem fields, search attachments, or generate the image again merely to inspect it.
7. The Artifact path is a stable read_image reference, not a filesystem destination. If the user asks to place the image in the workspace or another user-visible folder, use the separately returned exact savedPath with the ordinary authorized file/command path. Never derive savedPath from the Artifact URI, and do not claim the image was exported until that separate operation succeeds.
8. A model with image-input capability may also receive the generated pixels transiently during the generation execution. visualInputDelivery uses stable values such as attachedDuringGeneration; it does not mean pixels remain attached in later model requests. In a later turn, call read_image with the exact returned path before claiming to have visually re-inspected the image.

Never send a URL, API key, model ID, provider ID, executable, raw base64, or data URL. Provider selection, credentials, model configuration, input encoding, execution identity, and Artifact storage are host responsibilities. Do not replace this tool with curl, a custom network request, or an ad-hoc script.

If the tool is unavailable or reports a configuration error, preserve that result and tell the user to review the image-generation Skill switch in Settings → Skills and the API/model configuration in Settings → Configuration → Image Generation; do not attempt a network fallback. Preserve failed, cancelled, and indeterminate outcomes exactly. Always inspect failure.retryable before considering another call. When it is false, do not retry it automatically; report the failure and follow the returned recovery guidance. When it is true, retry at most once and only when that still matches the user's intent. Never loop retries. An indeterminate result may represent a request that reached the provider and must never be retried automatically.

Product watermark

Treat a platform-added watermark such as “AI生成” as a product configuration result, not an image quality defect. Do not change that configuration or make the watermark invisible by cropping, covering, repainting, editing, or regenerating. If the user does not want it, direct them to turn off “添加水印” in “设置 → 配置 → 图片生成”, save, and generate again. This rule concerns the product's generated-image watermark, not third-party copyright watermarks or marks of unknown origin.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Tiga001
• Source: Tiga001/Captain_Who
• License: Apache-2.0
• Homepage: https://captainwhoagent.com

Install and usage instructions live in the source repository linked above.