Build an Accio Skill for Weekly Store Health Diagnosis (2026)

Core rule: Use uploadable store-health-weekly for weekly store-health diagnosis—copy only official diagnosis or Seller Central-visible metrics, overlay 24-hour overdue inquiry/RFQ queues, suggest P0 routes, and wait for human review. Never invent star ratings or Trade GMV; timezone Asia/Shanghai. For rating rules read the star encyclopedia; for RFQ quote bodies use K06.

1. Why a weekly store-health diagnosis Skill

Monday stand-ups fail when three people bring three stories: ops says stars slipped, sales says inquiries piled up, the owner asks why Trade GMV “moved again.” This Skill turns “is the store healthy this week?” into an auditable sheet: copy only official diagnosis returns or Seller Central-visible values, overlay 24-hour overdue inquiry/RFQ queues, suggest P0 routes, and stop for human review before anyone dispatches work.

It is not the encyclopedia article on merchant star ratings and upgrades. That page explains rules, benefits, and how to think about upgrades. K07 forbids inventing ranking weights or “target stars” inside a weekly report. Copy what you can see; write unseen or non-GGS when you cannot.

It also sits beside K06 RFQ quotation pack, not on top of it. K06 locks RFQ facts and drafts an English quote for a human to send. K07 never inserts quote paragraphs, unit prices, or lead-time promises into a health report. If your task prompt says “and quote the overdue RFQs,” you attached the wrong Skill to the wrong Agent.

Shipping this as an Accio Work uploadable pack—not a private spreadsheet—keeps day shift and night shift on one Hard-rules sheet: timezone Asia/Shanghai, no invented stars or Trade GMV, no filling gaps with last week’s numbers when diagnosis fails, and no auto send or settings changes. New to Skills? Start with K01, then come back. Hub: Accio Tutorials.

Acceptance is blunt: the weekly output must not contain “we will drop half a star next week,” “conversion is roughly x–y%,” or “plug 120k Trade GMV from experience.” Those lines mean invention, not copying. The Skill’s value is auditability.

From a management angle this is a risk radar, not a growth-copy machine. Radars optimize for few misses and explainable false alarms; copy machines optimize for tone. If your team starts optimizing “how inspiring the weekly reads,” stop and return to consistency with Seller Central.

2. Download store-health-weekly

Open SKILL.md first. Read the YAML name/description and hard rules: report only diagnosis returns or back-office-visible values; write unseen when non-GGS or tools fail; timezone Asia/Shanghai; never invent star ratings, ranking weights, conversion rates, or Trade GMV. If your team only wants a traffic narrative without five-dimension diagnosis, that is alibaba-analysis-brief—excludes already say so; do not delete the brakes to force a prose generator.

The pack includes references/health-checklist.md and evals/ fixtures. The checklist is for humans; the fixtures regress “non-GGS must not invent stars” and “failed diagnosis must not invent Trade.” Small zip, discipline is the product.

Related downloads with their own tutorials: inquiry-triage-priority.zip (K05) and the K06 RFQ pack. The weekly pack does not draft inquiry replies or RFQ bodies—mixing those outputs blurs who approved external commitments.

3. Reading SKILL.md: When / Hard rules / timezone

Key excerpts match the download:

---
name: alibaba-store-health
description: "Alibaba.com seller store-health audit: call query_ggs_merchant_info and query_store_diagnosis_daily_odps via alibaba-global-store-diagnosis, copy only returned or Seller Central-visible metrics with Asia/Shanghai timestamps, overlay the 24-hour inquiry/RFQ overdue queue, and route P0. Use for daily or weekly health checks, overdue 24-hour queues, TA status, expiring certs. Write 未见 if the merchant is not GGS or the diagnosis tool fails. Do not invent ranking weights, conversion rates, or unseen scores."
version: "2.0.0"
enabled: true
triggers:
  - 店铺健康
  - 店铺诊断
  - 星等级
  - 待回复队列
  - 响应率
  - store health
  - 过期询盘
  - GGS
examples:
  - 今天店铺正不正常，先跑官方诊断再叠 24 小时队列
  - 我们不是 GGS，星等级是多少
  - 诊断工具失败了，Trade GMV 写多少
  - 列出已过 24 小时未回的询盘并派单
  - 只要经营日报不要诊断
excludes:
  - skill: alibaba-global-store-diagnosis
    when: 用户只要官方完整诊断报告/IFM 升级引导，不要本包的 24h 队列叠加层
  - skill: alibaba-analysis-brief
    when: 用户要经营日周报/流量转化，且不是 GGS 诊断维 + 24h 队列
  - skill: alibaba-inquiry-conversion
    when: 用户要起草某条询盘回复而不是盘点
  - skill: alibaba-rfq-quotation
    when: 用户要写报价正文
  - skill: alibaba-rfq-discovery
    when: 用户只要搜 RFQ / 报价历史，不要健康表
  - skill: alibaba-product-optimization
    when: 用户要改某个 listing
  - skill: alibaba-trade-assurance
    when: 用户只要 TA 三件套或拒付，不要整店盘点
metadata:
  openclaw:
    author: gd-alibaba-ops
    homepage: https://www.accio.com/work
    emoji: "🏥"
---

# Alibaba.com 店铺健康盘点（Store Health）

对内中文。只报官方诊断返回或后台可见原值。不虚构排名机制。不对买家写英文信。时区固定 Asia/Shanghai。

## Quick Demo

用户：「日盘。GGS 店铺。再把过 24 小时的询盘列出来。」

智体：

1. 经 `alibaba-global-store-diagnosis` 调 `query_ggs_merchant_info`（是否 GGS：hkgs/twgs/tp）和 `query_store_diagnosis_daily_odps`。
2. 抄 Star Rating / Buyer Traffic / Business Opportunities / Product / Trade 原值，每条带 `copied_at`。
3. 叠加 24h 过期队列（聊天分析 + RFQ quote-history）。规则 URL：https://seller.alibaba.com/learningcenter/content/detail/PX2FWJSX.htm
4. 非 GGS 或工具失败：整维 `未见` / `非GGS`，不发明星级或 Trade 数字。

## Capabilities

- GGS 判定与五维抄写（只抄，不解释未见的权重）。
- 非 GGS / 工具失败 → `未见`，官方诊断技能对该人群只给升级引导。
- 24h 询盘/RFQ 过期队列叠加（本包质量层，不是诊断公式）。
- P0 分诊到询盘/RFQ/商品/TA/用户本人。
- 可选 `alibaba-analysis-brief` 做非诊断经营简报。

## How It Works

1. **优先官方诊断**：`alibaba-global-store-diagnosis`。工具：
   - `query_ggs_merchant_info`
   - `query_store_diagnosis_daily_odps`
   维顺序：Star Rating / Buyer Traffic / Business Opportunities / Product / Trade。仅 GGS 付费（hkgs/twgs/tp）。非会员给升级引导，不诊断。
2. **经营简报**（非五维诊断）：`alibaba-analysis-brief`。
3. **叠加 24h 队列**（诊断不提供逐条询盘时由本技能补）：
   - 询盘到达/过窗：`alibaba-global-chat-analysis-pro` / `alibaba-chat-and-analysis`
   - RFQ 已报价历史：`alibaba-rfq-discovery` + `workctl workflow rfq quote-history`
4. 每条 `copied_metrics` 必须有 `copied_at`（Asia/Shanghai）和允许的 `source`。禁止 `invented=true`。
5. 断言 24h 时必须出现 Learning Center URL。不要发明 `workctl ta` 或询盘 workctl。

允许的 metric source：`alibaba-global-store-diagnosis`、`query_ggs_merchant_info`、`query_store_diagnosis_daily_odps`、`alibaba-analysis-brief`、`seller-central-visible`、`alibaba-global-chat-analysis-pro`、`alibaba-chat-and-analysis`、`alibaba-rfq-discovery`、`未见`。

## Agent Instructions

硬规则：

- 全部询盘 24 小时内回复。https://seller.alibaba.com/learningcenter/content/detail/PX2FWJSX.htm
- RFQ 尽量 24 小时；否则先收悉。https://seller.alibaba.com/businessblogs/quality-quotations-win-quality-deals-px7uo467
- 博客 90%+ / 100% 属性 / 开 TA：标注「博客指导，非 SLA」。
- 星等级：只抄可见等级、分数、提示。公式与权重未见就不写。
- TA：开通与履约以后台和 https://tradeassurance.alibaba.com/ 为准。Trade 维数字只能抄诊断返回。

逐步（`references/health-checklist.md`）：

1. 写明可见范围：完整后台 / 仅截图 / 仅用户口述。口述不当作后台原值。
2. 先跑 `query_ggs_merchant_info`。非 GGS → `ggs_status=非GGS`，Star Rating / Trade 写 `未见` 或 `非GGS`。
3. GGS 则跑 `query_store_diagnosis_daily_odps`。打不开/失败：`ggs_status=未见` 或维值 `未见`，不用上次值填。
4. 待回复队列：列出到达时间与是否已过 24 小时。过窗是风险，不是可改写的历史。
5. TA：已开通/未开通/未见；可见纠纷原文抄。改 TA 设置 → 用户本人。
6. 证书有效期 vs 后台展示。过期派 `alibaba-listing-quality`。
7. **停顿**：没有任何可见数据且无截图 → 「无法盘点」+ 需要打开的页面类型，不编健康结论。
8. P0：清询盘 → `alibaba-inquiry-conversion`；RFQ 超时 → `alibaba-rfq-quotation`（事实仍经 discovery）；证书/类目 → listing + 官方优化；拒付 → TA 技能。

输出契约：`ggs_status` `timezone=Asia/Shanghai` `copied_metrics` `overdue_queue` `p0_routes`。`invented_star_rating` / `invented_trade_gmv` 必须为假或缺省。

禁止：编造响应率、星等级、排名权重、转化率、Trade GMV；把 90%+ 写成强制 SLA；用过期盘点冒充当前；自动回复或自动改设置；建议下平台关 TA 省钱。

## Error Handling

| 情况 | 动作 |
| --- | --- |
| 非 GGS | `ggs_status=非GGS`；星级/Trade 未见；可提示走官方诊断的升级引导 |
| `query_*` 失败 | 整维 `未见`，不发明数字 |
| 指标名称看不清 | 抄截图可见字符，不翻译成未出现的官方指标名 |
| 出现 24h 字样 | 输出必须含 Learning Center URL |
| 用户只要官方长报告 | handoff `alibaba-global-store-diagnosis` |
| 用户只要经营日报 | `alibaba-analysis-brief` |
| timezone 不是 Asia/Shanghai / UTC+8 / CST-8 | 校验失败，改回操作者时区 |

## Related Accio Skills

- `alibaba-global-store-diagnosis`：GGS 诊断与升级引导。
- `alibaba-analysis-brief`：非诊断经营简报。
- `alibaba-global-chat-analysis-pro` / `alibaba-chat-and-analysis`：询盘队列。
- `alibaba-rfq-discovery`：RFQ quote-history。
- `alibaba-inquiry-conversion` / `alibaba-rfq-quotation` / `alibaba-listing-quality` / `alibaba-trade-assurance`：P0 执行层。

## 官方硬约束

- 全部询盘 24 小时内回复。[How to respond to buyer inquiries](https://seller.alibaba.com/learningcenter/content/detail/PX2FWJSX.htm)
- 尽量 24 小时内回复 RFQ，否则先 ack。[Quality quotations win quality deals](https://seller.alibaba.com/businessblogs/quality-quotations-win-quality-deals-px7uo467)
- Seller Blog 2026 90%+ / 属性 100% / 开 TA：**博客不是 SLA**。
- 官方诊断 `alibaba-global-store-diagnosis`：只分析 `query_ggs_merchant_info` / `query_store_diagnosis_daily_odps` 返回；空/错该维写未见；仅 GGS（hkgs/twgs/tp）可诊断。
- Trade（TA GMV、订单、发货率、NR 取消）只能抄诊断返回。
- TA 未知数字：https://tradeassurance.alibaba.com/

## 禁止事项

- 编造响应率、星等级、排名权重、转化率、TA GMV。
- 把 90%+ 写成强制 SLA。
- 用过期盘点冒充当前。
- 自动回复或自动改设置。
- 建议下平台或关 TA 省钱且无官方页依据。
- 写「预计下周掉星」。
- 发明 workctl / 诊断枚举。

When stays narrow: daily/weekly health, overdue 24h queues, TA status, expiring certs—not a general business-essay writer. Keep task prompts equally narrow; never “use this Skill to fix the store automatically.”

Hard rules are brakes: reply to inquiries within 24 hours (Learning Center URL required when asserting 24h); RFQ within 24h when possible else acknowledge; stars only from visible fields; Trade numbers only from diagnosis; blog 90%+ / 100% attributes / enable TA labeled “blog guidance, not SLA.” Banned lines include predicted star drops, invented workctl, and “disable TA to save money” without official pages.

Timezone is fixed to Asia/Shanghai (UTC+8). If the runtime or output timezone is not Asia/Shanghai / UTC+8 / CST-8, validation fails—reset to the operator timezone. copied_at, inquiry arrival, and 24h deadlines all use this zone. Mixing buyer-local time without labels creates false “still early” states.

Human review is part of the contract: P0 routes are suggestions, not auto-dispatch; changing TA settings, auto-replies, or auto product edits are forbidden. With no visible data and no screenshots, output “cannot audit” plus which pages to open—never invent a soothing health conclusion.

4. Workflow (diagram)

Four steps: (1) query_ggs_merchant_info for GGS (hkgs/twgs/tp); (2) if GGS, run query_store_diagnosis_daily_odps and copy Star / Traffic / Opportunities / Product / Trade in order; (3) overlay 24h inquiry and RFQ overdue queues when diagnosis does not list each thread; (4) propose P0 routes and stop for humans.

Non-GGS: ggs_status=non-GGS, stars and Trade as unseen/non-GGS; you may point to upgrade guidance, never “about 4.8 stars.” Tool failure: entire dimensions unseen—do not backfill last week. That is stale audit presented as current, which hard rules ban.

Allowed metric sources are whitelisted in SKILL.md. Anything outside the whitelist, or invented=true, fails review.

5. Copy board: five dimensions + 24h queue

Details live in references/health-checklist.md. Oral claims are notes, not copied_metrics. Certs: copy displayed names and visible expiry; expired → listing quality Skill; claimed-but-unheld certs are severe risk (remove display, never forge). TA: on/off/unseen; unknown numbers → https://tradeassurance.alibaba.com/ — no fee invention.

6. Boundaries vs star encyclopedia and K06

• /benefits encyclopedia — how stars are rated and upgraded;
• K07 this pack — what Seller Central/diagnosis shows this week + 24h overlay + P0 suggestions;
• K06 RFQ pack — English quotation draft after fact lock, human Send.

Failure modes: pasting “push to four stars” goals into the weekly conclusion; filling Trade when diagnosis returned empty; letting the health Agent “helpfully” draft RFQ quotes. The first two break audit; the third breaks send accountability.

Neighbors: queue triage → K05; single-inquiry drafts → K02; listing gaps → K03. Weekly health answers “what is hot store-wide,” not “how to reply to this thread.”

7. Upload, bind Agent, test prompts

1. Zip / upload / test-run basics: follow K01 (SKILL.md at folder root). Then upload under Capabilities → Skills (or current client label).
2. Bind to a store-health / ops assistant Agent—do not mix as the default with RFQ-quoting or inquiry-reply Agents;
3. Prompt example: Run a weekly store-health diagnosis with the store-health-weekly skill. Copy only diagnosis or Seller Central-visible metrics with Asia/Shanghai timestamps. Overlay 24-hour overdue inquiry/RFQ queues. Route P0 suggestions. Do not invent star ratings or Trade GMV. Do not change settings. Wait for human review.
4. Check for ggs_status, timezone=Asia/Shanghai, copied_metrics, overdue_queue, p0_routes; Learning Center URL when asserting 24h;
5. After human OK, clear queues via inquiry/RFQ execution Skills; certs via listing; chargebacks via TA helpers—the weekly Agent does not run appeals or auto-replies.

Pressure-test three realities: GGS with diagnosis up; non-GGS; diagnosis failure. Green-path-only tests miss hallucinations. For three weeks, compare model copies to a supervisor sheet; edit the dimension text that drifts—do not swap models first.

8. Worked examples

User: Weekly check. GGS store. Run official diagnosis, list inquiries/RFQs past 24h, give P0.

Pass: visibility scope stated; GGS status; five dims with copied_at+source; overdue rows with times; Learning Center URL if 24h asserted; P0 routes stop for review. Fail: experiential Trade GMV, predicted star drop, or auto TA edits.

User: We are not GGS. What is our star rating? What Trade GMV should we write?

Pass: non-GGS status; stars/Trade unseen; optional upgrade pointer; never “about 4.8 stars.” Fixtures fail_invent_stars / fail_invent_trade_gmv exist for this.

User: Diagnosis failed—just paste last week’s Trade GMV; the owner meeting starts now.

Pass: dimensions unseen; refuse backfill; meeting can carry a “data unavailable” page. Stale-as-current is explicitly banned.

9. Failure cases, human gates, owner dashboard

Human gate SOP: weekly sheets enter a “pending confirm” queue, not a blast chat; P0 clearers must sign; settings changes need a change ticket. Sample one week of rows against Seller Central; fix Skill text or permissions when drift clusters—do not blame “AI mood.”

Owner dashboard five columns: GGS/diagnosis success, count of unseen dims, 24h overdue count, P0 sign-off rate, human override count. Friday fifteen minutes: rules or roster? Rules → Hard rules/checklist; roster → who may confirm P0.

Treat “unseen” as discipline, not embarrassment. Each unseen state has an action: upgrade path, re-run diagnosis, add screenshots, open a named page—never “invent a number so the meeting continues.” Track unseen count on the cover; owners need exposure, not pretty fiction.

Localization table: metric name, allowed sources, unseen wording, default P0 owner, last calibration date, notes. No “experiential GMV” or “target star” columns—those invite invention. Calibration older than 14 days → yellow flag.

10. Two-week rollout and acceptance

Days 1–2: upload and run the three examples; discuss only whether Hard rules held. Days 3–5: blind-test one GGS and one non-GGS store against a supervisor sheet. Days 6–8: write deltas into an appendix (P0 roster, unseen phrasing). Days 9–10: open the five-column dashboard. Days 11–14: freeze rules, tune roster only.

Do not change description, model, and roster in the same week—you will not know what worked. Never ship a “loose daily” that allows estimates; daily reuse the same brakes.

1. Zip opens to store-health-weekly/SKILL.md;
2. Agent bind succeeds;
3. GGS / non-GGS / failure examples obey Hard rules;
4. Output contract complete; 24h assertions include Learning Center URL;
5. Timezone Asia/Shanghai; no invented stars/GMV; no auto settings;
6. New hires can explain the encyclopedia vs K07 vs K06 split.

Series naming K01–K09 exists so people know which brake lesson to open—not so one Agent loads nine packs and “does everything.” K08 Trade Assurance checklist and K09 sample follow-up stay separate execution layers.

Finance/compliance one-liner: Trade figures without diagnosis source must not enter internal operating meetings. Spot-check source fields; the cost of checking is low, the cost of fake GMV decisions is not.

11. Misread → correct reading

Official long diagnosis / IFM upgrade handoff goes to alibaba-global-store-diagnosis. Traffic narratives without GGS dims + 24h overlay go to analysis-brief. This pack is diagnosis copy + quality-queue overlay—not a universal ops ghostwriter.

Prompt verbs to avoid: beautify, estimate, predict star drop, auto-reply, change price now. Prefer: copy, list, mark unseen, suggest route, wait for human review. Wrong verbs drift the model into invention and execution.

12. Gold-standard output shape

A gold-standard weekly packet always includes: visibility scope; ggs_status; timezone=Asia/Shanghai; copied_metrics[] with value, copied_at, source; overdue_queue[] with arrival and breach; p0_routes[] with suggested Skill and human gate; explicit false/omitted invented flags. Optional: cert expiry rows, TA on/off/unseen, checklist ticks from health-checklist.md.

Red flags in review: any percentage that did not appear on screen; any GMV with source “experience”; any star number when ggs_status is non-GGS; any sentence that changes settings; any RFQ quote body; any missing Learning Center URL beside a 24h claim.

Keep raw diagnosis fingerprints (or failure codes) with the model output and the human-final sheet. Monthly, sample eight packets. If humans frequently overwrite “unseen” with oral numbers, stop broadcasting—your process is inventing offline.

13. FAQ