Slime Amazon Design Skill

---
name: amazon-ai-design-workflow-orchestrator
description: "Use when the user invokes 史莱姆亚马逊设计skill or the Amazon 主图A+ / 主图+A+ design skill, asks for Amazon main-image and A+ workflow execution, or needs project-stage control across product references, research, style sampling, prompts, image2 generation, QC, and export."
---

# 史莱姆亚马逊设计skill

Skill version: `1.0.0`


This is the stable `1.0.0` release and the default for ordinary Amazon main-image/A+ requests.

## Image model default

`LATEST_IMAGE_MODEL`: apply the `imagegen` model policy before generation. Resolve the newest stable model from current official image-generation documentation at the start of a run; as checked on `2026-10-15`, prefer `gpt-image-2.5-sunburst` for product fidelity and precise edits. Keep that selection for the run and recheck for a new run. `image2`, `provider: image2` and `IMAGE2_PROVENANCE_GATE` are compatibility identifiers, not a model-version pin; they allow a newer model on the same supported route.

Select the model through a supported execution parameter or verify a current first-party runtime guarantee. Prompt text cannot select a backend. If the required latest model cannot be selected or verified, report `blocked: model_selection_unavailable` before generation; record actual model `unknown` and verification `unverified`, and do not claim an upgrade. A latest-model preference does not authorize a new paid API route or spending. Preserve current product, reference, copy, scope and QC gates and all historical generation records.

The controller records target model, actual model when exposed, verification status and official source in the existing run notes, separately from provider. Before dispatch, inline the resolved model/route and verification state as execution instructions outside the five-field image packet and outside the image prompt. Fresh workers must receive these instructions without reading extra files; if they cannot execute the verified route, they return a standard `OTHER` failure with `model_selection_unavailable`. Do not dispatch an unverified route to satisfy a strict latest-model requirement.

## Goal

Act as the full-flow controller for Amazon main-image and A+ production.

Calling this skill for a new image set means `FULL_FLOW` by default. Directed edits to an approved existing baseline follow `SCOPED_REVISION` below. Use `QUICK_TASK` only when the user explicitly asks for a sample, test image, first-screen test, or an ordinary 1-3 image task. A quick task stays in the current task and does not count as full-flow progress.

`FULL_FLOW_EXCLUSIVE_CONTROL`: once a request enters `FULL_FLOW`, generic brainstorming, visual-companion pages, HTML/CSS mockups, and nearby helper workflows may not replace, pause, or relabel the state machine unless the user explicitly requests that separate artifact. “先出样” under time pressure does not silently convert an already-invoked full run into `QUICK_TASK`.

The controller must:

- inspect the project before asking questions;
- ask only for missing production-scope fields;
- protect product truth and local details before style;
- generate at least three white-background reference images from the photographed product images at the beginning of the workflow;
- resolve a dynamic Product Facts Gate before big-brand research, style samples, or full production, asking only about remaining blocking facts;
- use the lightest valid reference route;
- research competitors and current big-brand cases for every product;
- find references by visual/layout purpose, reuse explicitly liked favorites, and refresh rejected candidates with minimal user input;
- compare three styles with one representative image per style;
- carry product-exposure role checks into the compact production context and formal production planning;
- continue selected-style formal production in the same task by default;
- generate every approved raw slot before centralized finished-image QC;
- use bounded subagent parallelism for independent research, style, planning, prompt, generation, and QC work when subagents are available;
- keep the controller as the only writer of canonical facts, decisions, state, and final manifests;
- create worker documents only when a worker/subagent is actually dispatched;