# AI prompts for ecommerce: 30 ready prompts for orders, stock and SEO

AI prompts for ecommerce: 30 ready prompts for an AI agent in your online store, each with the key it needs, for orders, stock, prices, SEO and email.

Source: https://cloudcart.com/en/blog/ai-prompts-online-store · CloudCart · 2026-11-06

In short: an AI prompt is the text instruction, question or task you give an AI model such as ChatGPT, Claude or Gemini. Connected to CloudCart with a personal access token, an AI agent carries out the prompt inside the store itself: prices, stock, discounts, SEO, emails and reports. Below are 30 ready prompts for ecommerce, each with the key (scope) it needs.

## What is an AI prompt, and how does it work in an online store?

An AI prompt is the instruction you give an AI model: a question, a task or a rule. Sent to a chat window, a prompt returns text that you copy by hand. Sent to an AI agent, a prompt leads to an action: the agent finds the data, shows you what it is about to do and, once you confirm, does it.

An AI agent is a program such as Claude Code, ChatGPT Codex, Gemini CLI, Cursor or VS Code with Copilot that runs on your computer and has access to your store. In CloudCart, the ecommerce platform the agent connects to through the Admin GraphQL API, it has more than 1,100 operations to work with: products, variants, stock, orders, customers, discounts, marketing and settings. Its permissions are those of the key you created in the admin, and nothing more.

|  | Prompt in a chat window | Prompt to an AI agent in CloudCart |

| What you get | Text to copy | A change in the store, or a report from your real data |

| Does it see the store? | No, only what you paste in | Yes, within the limits of the key |

| Bulk change | Not possible | "Raise the prices of brand X by 8%" for every product at once |

| Check before a change | Nothing to check | With the CloudCart skills it shows the count and the first 5 products, then waits for your yes |

| Numbers in the answer | General, from the model's training | Your orders, sales and visits |

Using AI in the store starts with 4 steps that take a few minutes:

- Create the key. In the admin open Settings → PAT Tokens (or the user menu at the bottom left → PAT Tokens) and create a token with the Read Only level to start.

- Copy the instruction. The "Connect your store to an AI agent" card on the dashboard has a "Copy the sentence" button that copies one sentence in English.

- Paste it into the assistant. The agent asks which language to work in and what the store sells, checks your computer, installs the CloudCart tools and asks for the token and the store address.

- Write your prompts. Start with a read-only question, for example prompt 1 or 28, and give write access only once you have seen how it works.

The prompts below work the same way in Claude Code, ChatGPT Codex, Gemini CLI, Cursor and VS Code. How each assistant connects is on the pages for [Claude Code](https://cloudcart.com/en/offer/ai-claude-code), [ChatGPT Codex](https://cloudcart.com/en/offer/ai-chatgpt-codex), [Gemini CLI](https://cloudcart.com/en/offer/ai-gemini-cli), [Cursor](https://cloudcart.com/en/offer/ai-cursor) and [VS Code](https://cloudcart.com/en/offer/ai-vs-code), and the overview is on the [CloudCart AI integration](https://cloudcart.com/en/offer/ai) page.

## How to write an AI prompt the agent gets right: the 5 parts

An AI agent makes the fewest mistakes when the prompt says exactly which products, what action, by which rule, how to check and what to leave alone. A missing part is the most common cause of surprises, because the agent then decides on its own what you meant.

| Part | The question it answers | Example |

| 1. Scope | Which products, orders or period? | "the variants of brand Lumio in the Lamps category" |

| 2. Action | What exactly should it do? | "raise the price" |

| 3. Rule | Which formula, with which limits? | "by 8%, round to .90, never above €200" |

| 4. Check | What should it show you before the change? | "show me the first 5 before and after and wait for my yes" |

| 5. Boundary | What must it not touch? | "leave products with an active discount alone" |

The prompt framework most guides recommend for chat has five elements: role, task, context, format and constraints. With an AI agent, the role and the general context are written once in the standing instructions (see below), and the agent reads the product context from the store itself. What stays in the prompt is the task (scope and action), the rule, the check as the format of the answer, and the boundary as the constraint.

Put together, the prompt looks like this:

Raise the prices of all variants of brand Lumio in the Lamps category by 8%.
Round to .90 and never go above €200. Leave products with an active discount alone.
First show me the count and the first 5 products with old and new price,
then wait for my "yes". At the end, list everything that did not go through.
Three habits save the most time. Write numbers with their unit ("€15", "15%", "until Sunday 23:59"). Name products, categories and brands exactly as they appear in the store. For a question or report add "read only, change nothing", and for a change always add "show me first".

### Standing instructions: the store's rules, written once

Rules that apply to every task are written once. In Claude Code this is the CLAUDE.md file in the folder you start the agent from, and in Codex it is AGENTS.md. The agent reads them at the start of every conversation:

Write in English and address customers as "you". Store prices include VAT.
Save new products as drafts only. Do not touch the Sale category
without my confirmation. Descriptions are 3 sentences, using only product facts.
Before any change to more than 1 product, show me the plan.

> Store run by an AI assistant One sentence to connect, and you decide the permissions CloudCart charges nothing for the connection; you pay the assistant's subscription to its provider. The assistant installs the tools itself, and with the personal access token you decide whether it only reads or can also make changes. How the AI assistant connects →

## Which key (scope) does the agent need for each prompt?

The agent's key is a personal access token (PAT) that only the store owner creates under Settings → PAT Tokens, up to 100 tokens per owner. It starts with cc_pat_, is shown once and has one of three levels: Read Only, Full Access or Custom Permissions by area.

| Work | Scope under Custom Permissions | Read only |

| Orders | orders | orders:read |

| Products, categories, properties, brands | products | products:read |

| Product stock | inventory | No separate read option |

| Customers and subscribers | customers | customers:read |

| Discounts and promo codes | discounts | No separate read option |

| Blog | blog | blog:read |

| Reports, campaigns and other marketing, settings, apps | settings | settings:read |

For questions and reports the simplest and safest choice is the Read Only level: the agent reads everything and cannot change anything. For changes, give Read & Write only on the areas the task needs. Inventory and Discounts only have full access to the area, so an agent that reads them under Custom Permissions needs inventory or discounts. A key can also have an expiry date and up to 20 IP addresses it works from, and switching it off stops the agent at once.

Important: the key acts on behalf of the store owner, with every permission its level allows. Give each agent and each recurring task its own key, and watch the "last used" column in the token list.

## 5 AI prompts for orders

Order prompts save the most time on the daily check and on bulk status changes. Reading needs orders:read, and changing a status needs orders.

| Prompt for the agent | What the agent does | Key |

| 1. "Show the orders from the last 24 hours that are not fulfilled yet: number, total, payment method and courier. Read only." | The morning packing list in one table | orders:read |

| 2. "Which unpaid orders have been pending for more than 3 days? Sort them by total." | Finds the orders to chase before stock is held up | orders:read |

| 3. "Mark the orders with these numbers as fulfilled and tell me which ones failed and why." | Starts one background job and reports what was skipped | orders |

| 4. "Which customers ordered for the first time this month, and what is their average order compared with returning customers?" | Compares new and repeat customers | orders:read + customers:read |

| 5. "Make a table of last month's orders by country and by courier, and show where cancellations and returns are highest." | Shows where you lose parcels and money, market by market | orders:read |

The platform enforces its status rules for the agent too. An order becomes "Completed" only if it is paid or fulfilled, and a paid or completed order cannot become "Cancelled". The background job for a status change returns the list of orders the rules refused, and the agent passes it on to you. Every status change is written to the order history.

## 4 AI prompts for stock and inventory

The most useful stock prompt combines stock level and sales speed, so you know what will run out before it does. The Products → Inventory screen lists only products with stock tracking switched on, and the agent counts by the same rule.

| Prompt for the agent | What the agent does | Key |

| 6. "Which products have fewer than 5 units and sold at least 3 times in the last 14 days?" | A list of best sellers that are running out | Read Only |

| 7. "Add the stock from this delivery using this table of SKUs and quantities. Show me old and new stock first." | Adds the units to each variant instead of overwriting them | inventory |

| 8. "Hide all out-of-stock products in the Summer Dresses category without deleting them." | A bulk change of the "hidden" flag | products |

| 9. "Make a supplier order list: products that will run out within 10 days at the current pace, with the quantity for 30 days." | A table ready for the supplier | Read Only |

On the Inventory screen a manual change never takes the quantity below 0, and every save refreshes search and the storefront cache. That is why for prompt 7 you should always ask for old and new stock before the save and check that no variant ends up with a negative number.

## 5 AI prompts for prices and promotions

A pricing prompt is most valuable when the rule differs per product: in the admin, the bulk "Change price" action gives every selected variant the same new price, while the agent applies the formula to each one. Variant prices need products, and discounts and promo codes need discounts.

| Prompt for the agent | What the agent does | Key |

| 10. "Raise the prices of all variants of brand X by 8% and round to .90. Show the first 5 before and after." | A new price for each variant by the formula | products |

| 11. "Create promo code BLACKFRIDAY for 15% off the Winter category from Friday 00:00 to Monday 23:59, once per customer." | A promo code limited by category, dates and uses per customer | discounts |

| 12. "Create a quantity discount for House Blend Coffee: €11 each for 2, €10 each for 4." | "Buy more, pay less" for one product (up to 12 tiers) | discounts |

| 13. "Which products have stock but no sale in 60 days? Suggest 20% off until Sunday for those only and wait for my yes." | Clears slow stock without discounting everything | Read Only, then discounts |

| 14. "Show all active discounts that have expired or have not been used in 90 days, and switch them off after my yes." | Cleans out old promotions that confuse customers | discounts |

Three CloudCart discount rules apply to the agent too. A store can have only one countdown discount, even a disabled one, so the old one is deleted to make a new one. A discount without a code (percentage, amount, fixed price or shipping) can be switched on or off at most once every 10 minutes. By default a promo code does not apply if the cart already holds a discounted product. All 8 discount types and how to create them by hand are in the guide to [promo codes and discounts in CloudCart](https://cloudcart.com/en/blog/discounts-promo-codes), and Black Friday ideas are in the article on [10 offers for Black Friday and Christmas](https://cloudcart.com/en/blog/promo-codes-discounts-online-store).

Important: anywhere in the EU, an announced price reduction must show the prior price, which is the lowest price you applied in at least the 30 days before the reduction (Article 6a of the Price Indication Directive, added by the [Omnibus Directive (EU) 2019/2161](https://eur-lex.europa.eu/eli/dir/2019/2161/oj/eng), applicable since 28 May 2022). Add to prompt 10 and 13: "check that the crossed-out price is the lowest of the last 30 days".

> 0% commission The promotion lifts revenue, and CloudCart takes no share of it If the Black Friday code doubles your sales, CloudCart deducts no percentage from any order: you pay a fixed monthly plan on every plan. How 0% commission works →

## 5 AI prompts for the catalog, SEO and blog

Catalog and SEO prompts handle the work that gets postponed most often because it takes hours: titles, descriptions, properties and category texts for hundreds of products. The catalog needs products, and the blog needs blog.

| Prompt for the agent | What the agent does | Key |

| 15. "Write an SEO title up to 60 characters and a meta description up to 155 characters for every product in the Mugs category. Start with the product type. Show the first 10." | Writes and tracks the length for you | products |

| 16. "Which products have a description under 50 words? Write new ones only from the product data: material, size, care. Nothing invented." | Complete descriptions without false claims | products |

| 17. "Create a Material property with the values cotton, linen and viscose for the Dresses category, enable it as a filter and fill it in from the descriptions." | A category filter that helps customers narrow the choice | products |

| 18. "Which products have no brand, SKU, barcode or image? Give me a table by category." | Finds products that go out incomplete in Google Shopping and price comparison feeds | Read Only |

| 19. "Write a post 'How to choose a coffee mug' featuring our 5 best-selling models and save it to the blog as inactive." | A post built from real products that you review before publishing | blog + products:read |

The API's bulk operations cover flags, categories, brands, labels, stock and variant prices. The agent changes SEO titles and descriptions product by product. At a limit of 300 requests per minute, 500 products take a few minutes. If you sell in several languages, ask for each language in a separate prompt and name the language in it. What else your pages need to rank is in the [SEO checklist for an online store](https://cloudcart.com/en/blog/online-store-seo), and how AI search engines choose whom to cite is in the article on [GEO for ChatGPT, Gemini and AI Mode](https://cloudcart.com/en/blog/geo-ai-search-online-store).

> AURA The agent's descriptions become Aura Chat's answers The complete descriptions and properties the agent writes are the data Aura Chat answers customers from, and the data Aura Search uses to find the product even when the customer misspells it. What AURA does →

## 4 AI prompts for email marketing in Marketing Suite

The AI agent prepares an email campaign from segment to message, and sending stays with you. Marketing Suite is the Marketing → Marketing suite screen in the admin, where you see the results of email, SMS, Viber and Web Push. Customers and subscribers need customers, and campaigns, as part of marketing, need settings.

| Prompt for the agent | What the agent does | Key |

| 20. "Create a segment of customers with at least 2 orders who have not ordered in 90 days and have agreed to receive emails." | The segment fills in the background | customers + settings |

| 21. "Create a draft campaign for this segment with last month's 3 best sellers and a 10% code. Do not send it." | A draft with real products, images and the code | settings + discounts |

| 22. "Which products are left most often in abandoned carts? Write the text for the cart recovery email with them." | Copy for exactly the products customers leave behind | Read Only |

| 23. "How many opens, clicks, orders and revenue did each campaign bring in the last 30 days? Which one should we repeat?" | Compares campaigns by orders and revenue | Read Only |

Before sending, check the copy and recipients under Marketing → Campaigns. Unsubscribed contacts are skipped automatically. Under EU rules, marketing email needs the recipient's prior consent, with one exception: you may email your own customers about similar products if they could object when you collected the address and can object in every message (Article 13 of the [ePrivacy Directive 2002/58/EC](https://eur-lex.europa.eu/eli/dir/2002/58/oj/eng), applicable since 31 October 2003). That is why prompt 20 asks only for customers who agreed. How to build the campaign by hand in 5 boxes is in the guide to [a newsletter and email campaign in CloudCart](https://cloudcart.com/en/blog/email-campaign-cloudcart-step-by-step), and which 7 automated emails sell the most is in the article on [email marketing for an online store](https://cloudcart.com/en/blog/email-marketing-online-store).

## 4 AI prompts for Meta, Google Ads, TikTok and Search Console

For advertising, the AI agent does two things: it writes ads from your real products and sales, and it compares ad platform spend with store sales. It reads CloudCart data on its own. Data from Meta, Google Ads, TikTok and Search Console is not in the store, so you hand it over as a CSV export.

| Channel | Prompt for the agent | Key |

| Meta (Facebook and Instagram) | 24. "Write 3 Instagram ad variants for the 3 best sellers of the last 30 days: the benefit in the first line, the price, one call to action." | Read Only |

| Google Ads | 25. "Here is the Google Ads export. Compare spend per campaign with the Google Ads sales in the 'Sales by traffic source' report and work out the ROAS of each." | Read Only |

| TikTok | 26. "Which products have many visits but few sales? Pick 3 for TikTok and give an idea for a 15-second video for each." | Read Only |

| Google Search Console | 27. "Here is the Search Console export. Which product pages have many impressions and a CTR under 1%? Suggest new SEO titles and save them after my yes." | Read Only, then products |

Comparing with ad spend only makes sense if purchases are tracked. The pixels are set up once with the guide to [Meta Pixel and TikTok Pixel in CloudCart](https://cloudcart.com/en/blog/facebook-instagram-tiktok). The strategy for each channel is in the articles on [Meta Ads](https://cloudcart.com/en/blog/meta-ads-online-store), [Google Ads](https://cloudcart.com/en/blog/google-ads-online-store), [TikTok Shop and TikTok ads](https://cloudcart.com/en/blog/tiktok-shop-tiktok-ads) and [Google Search Console for an online store](https://cloudcart.com/en/blog/google-search-console-online-store).

## 3 AI prompts for metrics and the mobile store

The metrics that say the most about an online store are revenue, order count, average order value, conversion rate, abandoned carts, sales by source and the mobile share. The AI agent reads them straight from the CloudCart reports, without opening the Analytics screen and without a spreadsheet export.

| Prompt for the agent | What you get | Key |

| 28. "Make a report for the last 7 days against the previous 7: revenue, orders, average order value, conversion rate, abandoned carts, sales by source. Flag every number that fell by more than 10% and suggest one action for each." | The weekly report in one prompt | Read Only |

| 29. "Visits by device type and the cart funnel for Mobile and Desktop over the last 90 days. Where does mobile lose more customers?" | Whether mobile brings orders or only visits | Read Only |

| 30. "Find products without an image or with a name over 70 characters that display badly on a phone, and suggest short names." | A fix list for the mobile storefront | Read Only, then products |

The mobile store decides sales: in Europe, 54.45% of web page views in October came from phones, 43.46% from desktop and 2.09% from tablets ([StatCounter, Europe](https://gs.statcounter.com/platform-market-share/desktop-mobile-tablet/europe)). If prompt 29 shows many mobile visits and a lower conversion rate than desktop, start with the cart and phone speed using the checks in the article on [the mobile version of an online store](https://cloudcart.com/en/blog/mobile-version-online-store). How prompt 28 becomes a report every morning is in the article on [a daily store report from an AI agent](https://cloudcart.com/en/blog/ai-agent-daily-store-report). The formulas and benchmarks for each metric are in [the 10 KPIs an AI agent calculates for you](https://cloudcart.com/en/blog/ecommerce-kpi-metrics-ai-agent).

## How to check what the agent changed

You check in three places: the plan the agent shows before the change, the report it gives after it, and the product or order history in the admin. Changes made through the API have no general "undo" button, the same as changes made in the admin, so the check comes first and the way back is kept separately.

- Ask for a plan. With the CloudCart skills (in Claude Code, Cursor, Gemini CLI and VS Code) the agent shows the count and the first 5 when 10 or more products are affected, then waits for your yes. Before deleting it always asks. In Codex the skills are not installed, so write "show me the plan first" in every task there.

- Keep the old values. For 100 or more rows the agent first saves the current values to a CSV file and runs a dry run. Above 1,000 rows that file is mandatory.

- Ask for a report at the end. "Give me a list of everything changed and everything that failed." If one product fails, the rest carry on and the failures are listed.

- Open the product history. In the product list, the Change log icon shows for each product which field changed, the old and new value, and when.

- Open the order history. Every status change is written as a line in the order history.

- Ask about discounts. Discounts have no change history, so ask the agent for a list of the discounts it created or changed.

- Switch on backups. Settings → Backup & Restore keeps a daily copy of the whole store. The feature is on every plan, and the service is switched on with a separate subscription on the same screen.

## Do the prompts work in German, Italian, Polish or other languages?

Yes. When you connect, the agent asks which language you want to work in, and from then on it understands and answers in that language. Only the connection sentence is in English, because it is a command to the assistant. Write the way you speak, but name categories and brands exactly as they appear in the store.

If you run a separate store for each country, each with its own language and currency, every store gets its own token, and you tell the agent which store a prompt is for. A prompt such as "write the description in German for the .de store" then lands in the right place. How languages, currencies and separate stores per country are set up is in the guide to [multilingual and multi-currency stores](https://cloudcart.com/en/blog/multilanguage-multicurrency-multistore). How ChatGPT and Claude connect step by step is in the articles [online store with ChatGPT and Codex](https://cloudcart.com/en/blog/online-store-with-chatgpt) and [how to run your online store with Claude Code](https://cloudcart.com/en/blog/online-store-with-claude).

## Why run the store through an AI agent: how much it gives back

Running the store through an AI agent pays off most when a task spans several screens, affects many products or has to happen today. In the worked example for a store with 400 products, the agent saves about 19 hours of work a month, and two price lists take 0.5 hours instead of 4.4 hours by hand (the calculation is in the article [AI agent for an online store: how many hours and how much money it saves](https://cloudcart.com/en/blog/ai-agent-online-store-why)).

| Task | By hand in the admin | With one prompt to the agent |

| Which best sellers are running out | Three screens (report, products, inventory) plus a spreadsheet | Prompt 6 |

| New price by formula for 300 variants | Export, formulas, import | Prompt 10 |

| SEO titles for 500 products | Product by product | Prompt 15, a few minutes |

| Weekly report with conclusions | Each report separately, compared by hand | Prompt 28 |

| Campaign for inactive customers | Segment, template, products and code one by one | Prompts 20 and 21 |

The second gain is a new task without a new workflow. Rule-based automation only does what it was set up for, while the agent carries out a task you thought of this morning. The third is that ads, emails and descriptions are written from your own products, prices and sales, and before and after every change you see revenue, conversion rate and average order value.

> New store Try the 30 prompts in a real store for €1/month New stores pay €1/month for the first 3 months on every plan, with no code. That is enough time to see which prompts bring orders in your store. Create a store for €1 →

## Prompt mistakes that cost time

- A change with no scope. "Lower the prices" with no category, brand or list leaves the choice to the agent. Always say which products.

- Full Access for a report. Read Only is enough for questions and reports. Give Full Access once you know how the agent works.

- One long task instead of three short ones. Prices, descriptions and a campaign in one message are hard to check. Split them and start a new conversation between them.

- A formula with no limit. "+10%" without a ceiling or rounding gives prices like €19.47. Add "round to .90" and "never above €X".

- Text with no source. "Write a description" without "only from the product data" produces features the product does not have.

- Publishing without review. Save new products as drafts, and campaigns as drafts you send yourself.

Questions about connecting the agent or about a prompt that did not do what you expected: support@cloudcart.com.

## Frequently asked questions

**What is an AI prompt?**

An AI prompt is the text instruction you give an AI model such as ChatGPT, Claude or Gemini: a question, a task or a rule. In a chat window it returns text, while an AI agent connected to the store turns it into an action, for example a price change or a report from your orders.

**What are the best AI prompts for ecommerce?**

The most useful prompts combine data from several screens: which best sellers are running out, which products have not sold in 60 days, and the weekly report that flags drops above 10%. Next come bulk tasks such as SEO titles and prices by formula, which take hours by hand.

**How do I write a prompt so the AI does not make mistakes?**

State the scope, action, rule, check and boundary: which products, what to do, by which formula, what to show you first and what to leave alone. For a change add "show me first and wait for my yes", and for a report add "read only, change nothing".

**Can I write prompts in my own language?**

Yes. When you connect, the agent asks which language to work in and then understands and answers in it, whether that is German, Italian, Polish or another language. Only the connection sentence is in English, because it is a command to the assistant.

**How do I check what the AI agent changed?**

Ask for a plan before the change and a list of what changed after it. In the admin, each product's change log shows old and new values, and the order history records every status change. Discounts have no history, so ask the agent which ones it created.

**Which key should I give the AI agent?**

For questions and reports, create a personal access token with the Read Only level. For changes, choose Custom Permissions with only the areas you need, for example products for prices and descriptions or discounts for promo codes. A key can have an expiry date and an IP limit, and you switch it off at once under Settings → PAT Tokens.

**Can I use these prompts in ChatGPT?**

Yes, in Codex, OpenAI's agent in ChatGPT, once you connect it to CloudCart. The prompts are the same for Claude Code, ChatGPT Codex, Gemini CLI, Cursor and VS Code. In a regular chat with no connection to the store you only get text to copy.