# ecom-suite-forge
[English](README.md) · [简体中文](README.zh-CN.md)
[](https://skills.sh/linbei0/ecom-suite-forge)
[](LICENSE)
Turn one set of viral e-commerce product images (爆款套图) into a single reusable, product-agnostic **suite template (套图模版)**.
The skill reverse-engineers the *visual grammar* of a winning image set — shot roles, composition, camera, lighting, background, color system, text strategy — and rewrites it as placeholder-driven prompt templates. Swap in a different product later and the suite still works.
> [!NOTE]
> This is an [Agent Skill](https://agentskills.io). It forges templates only — it does not generate images and does not depend on any specific app, backend, or provider.
## What it does
Given a folder of 6–12 viral product images of the same product, the skill produces one `*.suite.json` file containing:
- an **L1/L2 category** and free-text `leaf` product name,
- a **conversion-ordered storyboard** mapping each source image onto a fixed 8-role funnel vocabulary (`HERO`, `PAIN_POINT`, `COMPARISON`, `SCENE`, `DETAIL`, `TRUST`, `VARIANT`, `CTA`),
- a **Campaign Style Lock** reused verbatim across every shot,
- one **English `promptTemplate` per shot** with hex colors, numeric product occupancy and a negative list.
At generation time the flow is:
```
user product photo (product truth)
+ suite template (N storyboard shots)
-> an agent rewrites each shot with the user's product facts
-> N final image prompts
-> a professional, high-CTR product image set
```
## Install
```bash
npx skills add linbei0/ecom-suite-forge
```
Install a specific agent or a global scope:
```bash
npx skills add linbei0/ecom-suite-forge -a claude-code -a opencode
npx skills add linbei0/ecom-suite-forge -g -y
```
Works with any Agent Skills–compatible agent (Claude Code, Cursor, Codex, OpenCode, Gemini CLI, and more).
<details>
<summary>Manual install</summary>
Copy this repository into your agent's skills directory, e.g.:
- project: `.agents/skills/ecom-suite-forge/`
- global: `~/.config/opencode/skills/ecom-suite-forge/` or `~/.claude/skills/ecom-suite-forge/`
`SKILL.md` must stay at the skill folder root; keep `references/` and `assets/` alongside it.
</details>
## When it triggers
Use it when you have a folder or batch of 爆款电商图片 and want to 拆解 / 提炼 / 复刻 them into a reusable 套图模版, or when you are building and expanding a 套图库.
> [!IMPORTANT]
> One suite = one main product. A suite needs multiple shots to be meaningful — the minimum useful set is about 5, 6–12 is ideal. This skill never invents certifications, ratings, sales counts, or efficacy claims; unverifiable proof becomes a `proof placeholder`.
## Output
A single file named `<l1-slug>-<leaf-slug>.suite.json`, for example:
```jsonc
{
"schemaVersion": 1,
"kind": "ecom.suite",
"id": "suite-hufugehu-jiemianru",
"name": "氨基酸温和不紧绷洁面乳款",
"category": { "l1": "护肤个护", "l2": "面部护理", "leaf": "氨基酸温和洁面乳", "leafKeywords": ["洁面乳", "氨基酸", "温和"] },
"styleLock": { "lockText": "Campaign Style Lock: ..." },
"shots": [
{ "shotId": "shot-01", "order": 1, "shotRole": "HERO", "promptTemplate": "Product photography of {product}. ..." }
],
"provenance": { "sourceKind": "viral-reference-set", "sourceImageCount": 8, "detached": true }
}
```
The full field contract lives in [`references/suite-schema.md`](references/suite-schema.md); a complete worked example is [`assets/suite-template.example.json`](assets/suite-template.example.json).
## How it works
The skill runs seven phases in order and gates the result before output:
1. **Intake** — confirm the whole set and the main product.
2. **Read & classify** — inspect every image, pick exact `L1`/`L2`, write a free-text `leaf`.
3. **Decompose each image** — fill the per-image analysis worksheet.
4. **Derive the Campaign Style Lock** — one whole-set visual contract.
5. **Write each shot's prompt template** — English, hex colors, numeric occupancy, negative list.
6. **De-identify + QA gate** — strip brand, logos, faces and source copy; all P0 checks must pass.
7. **Output** — write the suite JSON and report assumptions.
## Repository layout
```
SKILL.md # entry point: scope, iron rules, workflow
references/
category-taxonomy.md # built-in L1/L2 classification list
shot-playbook.md # per-shot purpose, composition, camera, lighting
prompt-contract.md # prompt iron rules + Campaign Style Lock
de-identification.md # what to strip from the source, and the audit
suite-schema.md # suite JSON schema (v1)
quality-checklist.md # P0/P1/P2 gate
assets/
analysis-worksheet.md # per-image analysis worksheet
suite-template.example.json # complete example suite
skills.sh.json # skills.sh repository page config
README.md # English guide (this file)
README.zh-CN.md # Simplified Chinese guide
```
## Core rules
- **Detach from the source.** Only reusable visual grammar survives; brand, model identity, logos, and source copy are stripped.
- **Order is a funnel**, not a pile of images: click → benefit → pain → proof → scene → detail → trust → variant → close.
- **Colors are hex**, never adjectives; product occupancy is a number; every prompt ends with a concrete negative list.
- **One style lock, reused verbatim** — the set stays consistent while angles, backgrounds and purposes vary.
- **No invented facts.** Unverifiable proof is left as a placeholder.
## License
Released under the [MIT License](LICENSE).