dsh-shop-assistant

中文 | English

For shop owners, CS leads, and operators — you do not need to write code.

You do not need to know English tool names. Follow the cases step by step.

In one minute

After you install this DeepSeek Harness (dsh) plugin, you can do three jobs in chat with plain language:

1. Batch bad-review replies — put an exported review spreadsheet in a folder; get many paste-ready replies that follow your return policy.
2. New listing copy from a competitor page — paste a public product URL; the assistant summarizes the page, then drafts titles, bullets, and FAQs.
3. Go / No-Go before listing — give cost, price, and 1–5 scores; a fixed formula computes profit and a recommendation (not a made-up guess).

This is not “just another chatbot.” Versus pasting into a web AI chat, you get whole-table handling, stable policy wording, reproducible math, and less copy-paste.

---

Install

1. Run DeepSeek Harness (e.g. npx @deepseek-ai/dsh web).
2. Install this plugin:

dsh plugin --profile web add dsh-shop-assistant
# or
dsh plugin --profile web add github:pengzhou267-ai/dsh-shop-assistant

3. Restart the Web UI or open a new session.
4. Pick a workspace folder (next section), then chat.

---

Before you start: where do files go?

What is the “workspace”?

It is the folder you select when you start dsh Web chat.
The assistant reliably reads tables and docs inside that folder only.

Suggested layout:

my-shop-files/
├── reviews.csv           ← your exported reviews
├── after-sales-policy.md ← your return rules
└── (optional) products.csv

Try without your own data first?

Copy samples from this package into the workspace:

File	Use
examples/reviews.csv	Fake reviews for case 1
examples/products.csv	Fake products
examples/score-inputs.csv	Numbers for scoring
kb/sample/售后政策.md	Sample return policy (edit before real use)

Header formats: examples/README.zh.md (Chinese; table headers are the same).

---

Case 1: Batch bad-review replies (daily)

How people usually do it

Copy reviews one by one from the seller console → paste into ChatGPT / DeepSeek web → re-explain return rules every time → paste replies back. Long threads blow up; wording drifts.

Prepare

1. Export reviews from Taobao / Pinduoduo / etc. Save as CSV UTF-8 if needed.
2. Prefer columns like: order id, rating, review text, date, SKU (see examples/reviews.csv).
3. Put the file in the workspace, e.g. reviews.csv.
4. Put return rules in after-sales-policy.md (start from kb/sample/售后政策.md).

Steps

1. Open dsh Web; set workspace to that folder.
2. Confirm you can see reviews.csv and the policy file.
3. Paste and send:

The workspace has reviews.csv and after-sales-policy.md (or 售后政策.md).

Please use the “read review spreadsheet” feature to open reviews.csv
(use the Taobao-style column mapping if headers look like a Taobao export).
Do not ask me to paste the table into chat.

Then:
1) Group bad reviews by reason (shipping delay, color mismatch, damage, size, …);
2) Write paste-ready replies for each group;
3) Strictly follow the policy file — no promises that are not written there.

(You may see tools like shop_csv_preview in the UI — you do not type those names yourself.)

What you get

Grouped, copy-paste replies keyed by reason / order id, aligned with your policy.

Compare

	Web AI chat	This plugin
Input	Paste into the dialog	Whole CSV in the workspace
Many rows	Context overflow	Whole-table pass
Policy	Re-typed every turn	Fixed policy file

---

Case 2: New listing copy (weekly / campaigns)

How people usually do it

Open competitor tabs → hand-copy titles → paste into an AI for polish. Slow; prices get wrong or invented.

Prepare

Copy a public product URL from the browser address bar (buyer-visible page, not a login-only seller console).

Steps

Send something like:

First, fetch information from this public product page (title, description summary, visible price clues).
Do not ask me to log into a seller console, and do not invent stock or promotions.

URL:
https://paste-a-real-public-product-url-here

Then follow the “new listing copy” flow and output:
1) 5 title options (with rough length);
2) five bullet points;
3) a detail-page outline;
4) 5–8 FAQs.
Our channel is Taobao. Core selling points: …

In plain words: the assistant summarizes the public page (shop_page_snapshot), then follows the built-in listing playbook (shop-listing). You only paste Chinese/English instructions and the link.

Compare

	Web AI chat	This plugin
Competitor info	You copy by hand	Paste public URL
Prices	Easy to invent	Prefer page price clues

---

Case 3: Pre-list profit check (weekly–monthly)

How people usually do it

Ask “cost 35, sell at 99 — how much do I make?” Numbers change every time.

Prepare

Field	Meaning	Example
cost	Unit cost	35
sell price	Your price	99
competitor price (optional)	Peers	109
demand / competition / ops / risk / timing	Scores 1–5	see prompt

See also examples/score-inputs.csv.

Steps

Please use the “profit scoring / product score” feature (fixed formula, no verbal guesses)
and explain in plain language: unit profit, margin rate, total score,
and whether to strongly recommend / caution / not recommend.

Cost 35, sell price 99, competitor 109;
demand 4, competition 3, ops difficulty 2, risk 2, timing 4.

You are asking the assistant to run the plugin formula (shop_product_score). You do not memorize the English name.

Same inputs → same outputs.

Compare

	Web AI chat	This plugin
Math	Improvised	Fixed formula
Repeat asks	Numbers may drift	Stable

---

Appendix

Shop-owner wording	What to say in chat	Internal name (optional)
Read CSV	“Use read-spreadsheet on xxx.csv”	shop_csv_preview
Fetch public page	“Fetch this public product page first”	shop_page_snapshot
Formula score	“Use profit scoring”	shop_product_score

Your own policy file

Copy kb/sample/售后政策.md, edit it, mention the path in the prompt. Advanced: set kbRelativeDir in the bundle config.

Contributing / license

See CONTRIBUTING.md and docs/EXTENDING.md. MIT.