Turn creative requirements into executable prompts, reference-image relationships, and size decisions. By default, the current Agent handles the rewriting; no separate prompt-enhancement model needs to be deployed. Reference-image tasks require visual understanding capability; text-only executors should first obtain an accurate image description.
First distinguish the task
- Prompt, rewriting, translation: Deliver copyable text and size suggestions, without submitting an image generation.
- Text-to-image: Describe the final visible image. Identity, count, text, action, color, and position are fixed requirements; only fill in the necessary composition and lighting details.
- Edit this image: First state what to change and what to change it to, then concisely lock down the unmodified content. When changing the pose, you cannot simultaneously require the pose to remain unchanged.
- Create a new image with a reference subject: First actually look at the image, clarify which image provides identity, clothing, or background, then describe the new pose and scene. Do not rely only on repeating character adjectives to maintain the same face.
Before writing the prompt, read rewriting.md. Preserve the user's creative intent; do not override specific requirements with generic beautifying words. Simple requirements do not need to be expanded into a long text.
Separate single images, series, and aspect ratio
A single image should arrange only one clear action. When a series' overall prompt contains multiple actions, split it into individually complete single-image prompts; for a same-character series, start separately from an identity reference that has passed review, to avoid accumulating the errors from the previous generation into each subsequent image. Only describe multiple panels when the user wants a collage.
Aspect ratio is an independent parameter. Explicit pixels or ratio take priority; when unspecified, suggest a ratio based on the composition and implement it in the downstream interface. Do not only say portrait in the text while submitting a square canvas. The same ratio can correspond to different pixel dimensions.
A structure that programs can connect to:
{"rewritten_prompt":"最终画面描述","wh_ratio":"9:16"}
{"rewritten_prompt":"具体编辑要求及保留条款","wh_ratio":"","ratio_follow":"<image1>"}
wh_ratio and ratio_follow are mutually exclusive. Only send rewritten_prompt as the text input to the model; do not also send the JSON wrapper or explanation into it. When the user only wants the prompt, delivering it as plain text is sufficient.
When generation is requested
Use the Qwen entry point specified by the user and already available, and continue with the current authorization. First read integration.md, and verify the actual interface: whether it supports reference images, multiple images, explicit width and height, and native Alpha. The model having a certain capability does not mean the current wrapper has integrated it.
This Skill does not preset server addresses or keys, nor does it download models by default. When there is no generation connection, prompts can still be delivered; do not describe 'request prepared' as 'generated'.
When the application needs to load the same versioned rules, read shared-runtime.md. The optional Node.js module only constructs planning messages, validates structure, and organizes quality feedback; it does not go online or execute generation; ordinary Agents do not need to run it.
The included scripts/prepare_workbench_request.py is only anoptional, offline, protocol-specific request builder for a particular workbench. It does not call models, does not perform prompt rewriting, and does not go online. Its 2048-pixel and single-reference-image limits apply only to that adapter, not to Qwen's model limits. See workbench-adapter.md。
Review and correction
Inspect the unretouched original image: target text, action, limb connections, identity, preserved areas, actual pixel dimensions; when transparency is needed, check Alpha and edges, rather than only looking at the file extension.
For each result, retain the original prompt, actual submitted text, reference images, generation settings, and result location. Record quality defects separately from aesthetic dislikes: multiple limbs do not mean a disliked sitting pose. When correcting, include the defect, target state, and content that must be preserved; keep the original image and the new version so the user can compare and accept. Stop when there is no further generation authorization; do not reroll indefinitely.
When the user wants an A/B comparison, fix the image model, size, steps, seed, and reference image, and change only the factor being tested. Compare experiments that adopt a new aspect ratio separately. Successful generation, a more aesthetically pleasing image, and strictly meeting the requirements are three different judgments.
When encountering text leakage, multiple panels, pose粘连, abnormal size, or transparent edge issues, read tested-behavior.md. These are observations under specific conditions and must not be expanded into general success rates.