git:20260620.1e9857f to git:20260910.9031ce8

- description: Use this skill when the user asks for a multi-image Amazon listing pack — typical surface forms include "做一套 Amazon 主图", "7 张套图", "Amazon listing 套图", "上架图包", "Amazon 7-pack", "listing image set". Generates the canonical 7-image set in PARALLEL — main shot, infographic, alternate angle, detail close-up, two lifestyle scenes, and scale reference — all sharing ONE img2img source so the same product carries across every angle. Open-ended prompts let img2img decide product-specific details (cup → handle close-up, wall art → back hook, rug → underside texture).

- ## When to use

- Match this skill on any user turn that asks for a coordinated multi-image Amazon

- listing set. Surface forms:

- - "做一套 Amazon 主图 / 7 张套图 / listing 套图 / 上架图包"

- - "Amazon 7-pack / listing image set / product photography pack"

- - "出一套上架图"

- Do NOT use:

- - Single-image requests → route to `image-prompt-sop`

- - Casual "画一张图" / "来张图" without "set / pack / 套图" wording — don't impose

- a 7-image flow on what the user wanted as one shot

- - Video requests → unrelated, route to `generate_video`

- ## Preflight: ATTACHMENT REQUIRED (hard gate)

- This skill is image-to-image. ALL 7 angles share ONE source image so the user's

- actual product carries across the set — that is the entire point of a "pack".

- Without a shared source, each angle is text-to-image from a generic "this

- product" prompt and you get 7 unrelated random products. The user's complaint

- "牛头不对马嘴" (results don't match the source) is exactly this failure mode.

- Before dispatching any `generate_image` calls, check that a system message

- labeled `[Canvas attachments for this turn]` is present in the conversation.

- If it is NOT, you MUST refuse and reply (match the user's language):

- - Chinese: "做这套图需要参考你的实际产品图。请先在 canvas 上选中一张产品图,点

- 右上 'Send to chat' 把它附上,然后重新发指令,我会按 7 个角度生成统一套图。"

- - English: "I need your actual product as a reference to generate this set.

- DO NOT call `generate_image` even once in the no-attachment case. The tool also

- hard-refuses pack calls (any call with `slot_index` set) that arrive without a

- source — so ignoring this preflight gets all 7 calls rejected at the tool layer

- with the error string spelled out anyway. Save the round-trip; refuse upfront.

- ## The 7-image template (Amazon canonical, product-class agnostic)

- Dispatch 7 `generate_image` calls IN PARALLEL within the same assistant turn.

- Each call has `n=1` (one distinct shot per angle, NOT n=4 batching the same prompt).

- The angles describe **what kind of shot** each slot is — but the prompts are

- deliberately open-ended. The model sees the source via img2img and decides the

- product-specific details (e.g. for slot #4 it picks the most distinctive detail

- itself: a cup's handle, a painting's back hook, a rug's underside texture, a

- chair's joint). You provide direction; img2img fills in the specifics.

- | # | Angle | Size | Prompt (open brief — keep it short, let img2img decide details) |

- | 1 | 主图-纯白背景 | `1:1` | Place this product centered on a pure white seamless studio background (RGB 255,255,255), product filling ~85% of the frame, soft even studio lighting, no shadow. Same product as the reference — colors, materials and design identical. Photorealistic, sharp focus. No text, no logo, no watermark, no props. |

- | 2 | 信息图-卖点标注 | `1:1` | Place this product on a subtle off-white / light gradient background. Add 3-5 thin arrows pointing at its most distinguishing features, each with a short, large, legible sans-serif label (1-2 words, layout only — don't rely on exact text rendering). Same product as the reference — colors, materials and design identical. Clean infographic style, photorealistic product. |

- | 3 | 另一角度 | `1:1` | Render this product from a notably different angle than the main shot (≈90° rotation or 3/4 view), revealing aspects not visible head-on. Same product as the reference — identical colors, materials and design; only the viewpoint changes. White seamless background, even studio lighting, photorealistic, sharp focus. |

- | 4 | 细节特写 | `1:1` | Macro close-up of this product's most distinguishing detail (model picks what best showcases craftsmanship — texture, mechanism, stitching, joint, edge, etc.). Same product as the reference — identical color and material, just magnified. Neutral background, bright even light, high detail, photorealistic. |

- | 5 | 场景图-自然光 | `1:1` | Place this product in a natural in-use lifestyle setting appropriate for its category (model decides the setting from what the product is). Same product as the reference — colors, materials and design identical. Mid-day natural light, shallow depth of field, editorial mood, 35mm, photorealistic. |

- | 6 | 场景图-暖光 | `1:1` | Place this product in a clearly different setting from the natural-light scene — a different room or style, not merely relit — under warm evening / golden-hour light, cozy mood. Same product as the reference — colors, materials and design identical. Shallow depth of field, 35mm, photorealistic. |

- | 7 | 尺寸对比 | `1:1` | Show this product alongside a clear scale reference (model picks an appropriate one — a hand for small items, a person or furniture for large items, a common object for mid-size). Same product as the reference — colors, materials and design identical. Neutral interior background, daylight, 35mm, photorealistic. No extra fingers, no distorted hands. |

- `size` 字段直接传比例字符串 (跟 frontend `IMAGE_EDIT_SIZES` 一致): `1:1` /

- `4:3` / `3:4` / `3:2` / `2:3` / `16:9` / `9:16` / `21:9` / `9:21` / `auto`。

- canvas provider (apimart) 原生收比例, 不要传 `1024x1024` 这种像素串。

- ## Per-call prompt construction

- Each prompt is a short img2img EDIT INSTRUCTION (1-2 sentences typically). The

- source attachment supplies the SUBJECT — you NEVER describe what the product IS

- ("modern canvas wall art", "ergonomic chair", "ceramic mug"). Every prompt

- opens with "this product" / "Place this product" / "Render this product"

- referring to the attached source. Fold in:

- 1. **The angle direction** from the table above (use the wording close to verbatim;

- the openness is intentional — img2img inherits the subject and visually

- appropriate details from the source).

- 2. **Identity-lock anchor (EVERY call)**: append a short clause pinning the

- product's identity to the source — `same product as the reference — colors,

- materials and design identical`. This is what keeps all 7 shots reading as

- ONE product instead of 7 near-misses (the classic "拼凑感"). Lock the

- product's *identity*, not its *pose*: the angle / scene / crop instruction

- already says what to vary, so phrase the anchor as identity ("same colors,

- materials, design"), NEVER as "identical image" — that can make img2img echo

- the source unchanged and ignore the re-angle / re-scene.

- 3. **Universal photo anchors**:

- - White / studio shots (#1, #2, #3, #4, #7): `photorealistic, sharp focus,

- high detail` (50mm-equivalent is implied for these — no need to spell out

- focal length on every shot)

- - Lifestyle shots (#5, #6): `35mm, shallow depth of field, editorial mood`

- 4. **Negative anchors when relevant**:

- - White-background shots: `no text, no logo, no watermark`

- - Main image (#1): also `no props`, product filling ~85% of the frame on

- pure white RGB(255,255,255) — Amazon main-image compliance

- - Shots with a person/hand (#7): `no extra fingers, no distorted hands`

- **Why short open-ended prompts**: gpt-image / Seedream-style img2img already

- sees the source. The more you over-prescribe ("hanging on a living room wall

- with sofa and plants softly blurred"), the more you bake in assumptions about

- product class — and if those assumptions don't match the source (the source is

- a cup, not wall art), the model EITHER ignores your prompt and does something

- random OR distorts the source to fit your prompt. Both bad. Keep the prompt

- short and let img2img inherit the right context.

- ## Tool call shape

- Within ONE assistant turn, dispatch all calls in parallel. **Each call MUST

- pass `image_urls` (the attachment URL from `[Canvas attachments for this turn]`),

- `label`, and `slot_index`** — `slot_index` (0-based) decides horizontal

- position; `label` titles the slot permanently (e.g. "1-主图-纯白背景" stays

- above the image after generation finishes); `image_urls` makes it img2img

- instead of text-to-image.

- The default 2-tool-call cap in the system prompt does NOT apply when this skill

- is active — `authorized-tool-calls: 7` in this skill's frontmatter lifts it.

- ### Label length — HARD LIMIT

- Each `label` MUST be **≤ 12 中文字 / 24 ASCII chars** total (including the

- `<index>-` prefix). The label renders above the placeholder rectangle in a

- fixed-width band on the canvas — anything longer overflows into the next slot

- visually. The frontend does NOT word-wrap labels, by design (a wrap would push

- the row down and break pack alignment).

- OK: `1-主图-纯白背景` (7 chars), `5-场景图-自然光` (8 chars), `Lifestyle scene 2` (16 ASCII).

- Too long → truncate to the essence:

- - ❌ `4-产品最有辨识度部位的特写细节` (15 chars) → ✂ `4-细节特写` (5 chars)

- - ❌ `Detailed close-up of the most distinguishing part` (49) → ✂ `Detail close-up` (15)

- If user explicitly asks for fewer than 7 ("3 张" / "5 张"), pick subset in

- priority order `1, 5, 2, 7, 3, 6, 4` (main → lifestyle → infographic → scale →

- alt-angle → second lifestyle → detail). Renumber `slot_index` to be 0..N-1

- contiguous so the canvas row stays gap-free.

- ## Anti-patterns

- - ❌ Running ANY `generate_image` call without `[Canvas attachments for this

- turn]` present — refuse upfront. The tool will hard-reject every pack call

- with `slot_index` set if it lands without a source image.

- - ❌ Describing the product subject in the prompt ("modern canvas wall art",

- "ergonomic chair", "ceramic mug") — the img2img source carries the subject;

- naming it in the prompt can override the source's actual shape.

- - ❌ Over-prescribing details that assume a product class — "hanging on a

- living room wall", "sawtooth hook on the back", "human silhouette next to it

- for scale". These bake wall-art assumptions into prompts that should work for

- cups / rugs / chairs / apparel too. Let img2img decide based on source.

- - ❌ Chaining the calls (waiting for #1 to finish before calling #2) — they're

- independent, dispatch all together in one assistant message.

- - ❌ Forgetting `image_urls` when an attachment IS present — every angle would

- look like a different product, the "set" loses meaning. (Tool has an auto-

- inject fallback for this, but rely on it as belt-and-suspenders only.)

- - ❌ Using `n=4` on a single call — that gives 4 variants of ONE angle, not 4

- distinct angles. Each angle needs its own call.

- - ❌ Putting brand names / trademarks in any prompt ("Apple-style mug"). The

- downstream IP infringement check tombstones the listing.

- - ❌ Promising specific text rendering in #2's infographic ("the label will say

- '16x24 inches'") — describe the layout, accept the model may render lorem-

- ipsum-ish text. User can request a re-do if labels matter.

- - ❌ Generating fewer than 7 without explicit user request to do so.

- ## Example invocation

- **User:** [attaches a product image via "Send to chat"] "帮我做一套 Amazon 主图"

- **Agent (single turn, 7 parallel tool_calls — all pass `image_urls`):**

- prompt="Place this product centered on a pure white seamless studio background (RGB 255,255,255), product filling ~85% of the frame, soft even studio lighting, no shadow. Same product as the reference — colors, materials and design identical. Photorealistic, sharp focus. No text, no logo, no watermark, no props.",

- prompt="Place this product on a subtle off-white / light gradient background. Add 3-5 thin arrows pointing at its most distinguishing features, each with a short, large, legible sans-serif label (1-2 words, layout only — don't rely on exact text rendering). Same product as the reference — colors, materials and design identical. Clean infographic style, photorealistic product.",

- prompt="Render this product from a notably different angle than the main shot (≈90° rotation or 3/4 view), revealing aspects not visible head-on. Same product as the reference — identical colors, materials and design; only the viewpoint changes. White seamless background, even studio lighting, photorealistic, sharp focus.",

- prompt="Macro close-up of this product's most distinguishing detail (model picks what best showcases craftsmanship — texture, mechanism, stitching, joint, edge, etc.). Same product as the reference — identical color and material, just magnified. Neutral background, bright even light, high detail, photorealistic.",

- prompt="Place this product in a natural in-use lifestyle setting appropriate for its category (model decides the setting from what the product is). Same product as the reference — colors, materials and design identical. Mid-day natural light, shallow depth of field, editorial mood, 35mm, photorealistic.",

- prompt="Place this product in a clearly different setting from the natural-light scene — a different room or style, not merely relit — under warm evening / golden-hour light, cozy mood. Same product as the reference — colors, materials and design identical. Shallow depth of field, 35mm, photorealistic.",

- prompt="Show this product alongside a clear scale reference (model picks an appropriate one — a hand for small items, a person or furniture for large items, a common object for mid-size). Same product as the reference — colors, materials and design identical. Neutral interior background, daylight, 35mm, photorealistic. No extra fingers, no distorted hands.",

- **Agent reply text**: "已开始生成 7 张套图 (主图 · 信息图 · 另一角度 · 细节特写 · 场景图×2 · 尺寸对比), 每张 15-30 秒, 会陆续出现在画布上。"

- **User (no attachment):** "做一套 Amazon 7 张主图"

- **Agent (single turn, NO tool calls — preflight refuses):**

- reply text: "做这套图需要参考你的实际产品图。请先在 canvas 上选中一张产品图,

- 点右上 'Send to chat' 把它附上,然后重新发指令,我会按 7 个角度生成统一套图。"