| name | image-prompt-sop |
| description | Use this skill whenever the user asks to generate, create, draw, or visualize an image on the canvas — including product photos, lifestyle scenes, mockups, marketing visuals, concept art, illustrations, or icons. The skill rewrites vague requests into high-quality prompts and picks the right size + count before calling `generate_image`. |
| allowed-tools | generate_image |
image-prompt-sop
When to use
Match this skill on any user turn that asks for a visual asset and would
naturally trigger generate_image. Typical surface forms:
- "画一个 / 帮我画 / 生成 / 来张 / 给我做个 / 出张图"
- "product photo / mockup / lifestyle shot / hero image / banner / concept art / illustration / icon"
- "show me what X looks like" / "visualize X"
Do NOT use this skill when the user wants a video (route to generate_video),
when they want feedback on an existing image (no tool call needed), or when
their request is so ambiguous that even the rewrite below would guess at
intent — in that case ask ONE clarifying question first.
The five-step rewrite
Before calling generate_image, mentally run through these five fields and
fold the answers into a single dense prompt string. Don't show the user the
breakdown — just call the tool with the finished prompt.
1. Subject + framing
Lead with what and how it's framed.
- Product: name it concretely ("a ceramic pour-over coffee dripper, matte
black, 1-cup size") not abstractly ("coffee equipment").
- Framing:
close-up macro / three-quarter product shot / top-down flat lay / wide environmental shot / eye-level lifestyle scene.
2. Lighting + mood
Pick ONE lighting setup and ONE mood word. Don't stack adjectives.
- Studio:
soft diffused softbox light from upper-left / bright even daylight, no shadow / dramatic single key light, deep shadows.
- Natural:
golden hour side light / overcast diffused daylight /
morning window light through sheer curtain.
- Mood:
clean / warm / moody / airy / editorial / cozy.
3. Background + surface
For e-commerce listings the default is pure white seamless unless the
user implies a scene.
- Pack shot:
pure white seamless background, no shadow
- Lifestyle:
on a [oak/marble/linen/concrete] surface, blurred [kitchen counter / living room / café] background, shallow depth of field
- Outdoor: name the locale concretely (
Mediterranean garden patio, not
outdoors).
4. Camera + render style
End the prompt with technical anchors that pin the output to "real photo"
vs "illustration" vs "render".
- Photo:
shot on Canon EOS R5, 85mm f/1.8, photorealistic, sharp focus on product, high detail
- Illustration:
flat vector illustration, limited palette (terracotta / sage / cream), thick outlines, no gradients
- 3D render:
octane render, soft global illumination, subsurface scattering on translucent parts
- Sketch:
loose pencil sketch on cream paper, single-weight line work, no shading
5. Negative anchors (only when needed)
Append explicit "no X" when a common failure mode is likely:
- People shots:
no extra fingers, no distorted hands, no warped face
- Text on packaging:
no text, no logo, no watermark, no typography
- Product packs:
no human hands holding the product
Skip this section if nothing obvious applies — empty anchors confuse some
providers.
Size selection
generate_image 收 aspect-ratio 字符串 (canvas provider 原生是 ratio):
1:1 / 4:3 / 3:4 / 3:2 / 2:3 / 16:9 / 9:16 / 21:9 / 9:21 / auto。
| Intent | Size | Why |
|---|
| Square hero / Amazon main image / Instagram post | 1:1 | Default. Pick this if unsure. |
| Portrait product detail / Pinterest pin / mobile splash | 2:3 | Vertical, more product real estate top-to-bottom. |
| Landscape banner / hero header / desktop wallpaper | 3:2 或 16:9 | Horizontal; 3:2 photo-classic, 16:9 cinematic. |
| 不确定要哪个 / 让 provider 看源图自动选 | auto | image-to-image 时跟着源图比例。 |
If the user says "for Amazon" → 1:1 (Amazon enforces square mains).
If the user says "for a banner / header / cover" → 3:2 或 16:9.
If the user says "for a story / reel / portrait" → 2:3 或 9:16.
Count (n) selection
| Situation | n |
|---|
| User wants a specific shot, knows what they want | 1 |
| User is exploring an idea, comparing styles | 2 |
| User is doing creative discovery, wants variety, said "a few options" / "几张" | 4 |
Never go above 4 (tool clamps anyway). Never chain multiple generate_image
calls in one turn — use n instead. The safety rules in the system prompt
enforce this.
Anti-patterns
- ❌ Calling
generate_image with the user's raw 5-word request. Always
rewrite through the five steps first.
- ❌ Stacking 8 lighting adjectives — picks one specific setup beats a
buzzword salad.
- ❌ "Beautiful, stunning, masterpiece, 8K, award-winning" filler. Models
trained post-2024 ignore these or actively bias toward kitsch.
- ❌ Asking the user to confirm the rewritten prompt before calling. Just
call the tool — they can ask for changes after seeing the result.
- ❌ Promising the image is "ready" or "here" in the reply — the tool is
async, the canvas updates when generation finishes. Say "started" /
"queued" / "should appear in 15–30 seconds".
Example rewrites
User: "画个咖啡杯"
Tool call: generate_image(prompt="A single ceramic espresso cup, off-white glaze with a tiny chip on the rim, on a worn oak café counter, soft window light from the left, shallow depth of field, photorealistic, shot on 50mm f/2.0, eye-level three-quarter view", size="1:1", n=1)
User: "给我做一个亚马逊主图,卖一个不锈钢搅拌器"
Tool call: generate_image(prompt="A stainless steel hand whisk, mirror-polished wire loops, ergonomic black silicone handle, centered product shot, pure white seamless background, soft even studio light, no shadow, photorealistic, sharp focus, high detail, no text, no logo, no watermark", size="1:1", n=1)
User: "几张极简风格的家居 mockup, 横版用做 banner"
Tool call: generate_image(prompt="Minimalist living room corner, a single linen-upholstered armchair beside a slim brass floor lamp, neutral plaster wall, light oak floor, morning daylight from an off-frame window, airy editorial mood, wide environmental shot, photorealistic, 35mm f/2.8", size="3:2", n=4)
