Use ChatGPT or Claude with your Shopify store. 30 things it can do, with the prompts.

30 things your store can do while you sleep. A Shopify AI agent is an AI assistant connected to your live Shopify store: it reads your real orders, products and inventory, and changes only what you approve. You type one sentence, it does the job, and shows its work.

See the 30 things

Prompts and the PDF are free with an email. Setup takes five minutes.

Done on real stores.

Four of the thirty. Store names are changed. The numbers are not.

Object 006123 products that had never sold once, found in one pull.See the card
Object 027A pixel had sent nine months of traffic data to someone else's ad account. Found, removed.See the card
Object 03117,722 sessions, zero orders, thirty days. The ad was optimising for the wrong event.See the card
Object 021108 product photos reframed to one look. 104 went live. Four never left the review folder.See the card

1 Ask your store.

Questions only, asked through Shopify's official connector for Claude or the Shopify plugin for ChatGPT. Nothing else to install, and you approve any write before it happens.

Read the setup and all ten cards in Series 1

Setup 1

Series 0, Set up

Connect Claude or ChatGPT to your store

Setup: No setup
Tools: Shopify connector for Claude, Shopify plugin for ChatGPT
Difficulty: easy
Wow

Before

You were copying numbers out of admin screens into a chat window.

After

The assistant reads the store itself and cites where each number came from.

Fill in

[nothing to fill in]: paste this as written. If the connector asks which store, pick the one whose admin URL you use.

The one rule

Grant the connector to a staff account with the permissions you would give a new employee, not the store owner login.

You should get back

The store name, the permission list, three numbers with their sources. If it cannot name the store, the connection did not complete.

Steps

In Claude, open Settings, then Connectors, and add Shopify. In ChatGPT, open Apps and add Shopify.
Sign in to the Shopify admin for the store you want connected.
Read the permission screen. It lists what the assistant can read and which writes need your confirmation. Approve it.
Back in the chat, run the prompt on this card.
If a number comes back with no source, disconnect and reconnect with the correct store.

The prompt

Object 001

Series 1, Ask your store

Sales by collection, this week and last year

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: easy
Wow

Before

Marlowe and Vine launched a candle collection on Monday and the founder was reading three admin screens to guess how it was doing.

After

One question returns the week's units and sales, the year-ago comparison and the product ranking, with the source named next to each figure and anything the report cannot scope to a collection called out as unavailable.

Fill in

[collection name]: the collection title exactly as it reads in your Shopify admin, for example the range you launched this week

The one rule

Ask where every number came from, because sessions and conversion rate are not reported per collection and a collection-level figure is usually the whole store.

You should get back

A summary table of six or so metrics with last year beside them and a Shopify report named on every row, then the products ranked by units sold. Sessions and conversion rate come back marked store-wide or unavailable rather than quietly scoped to the collection.

The prompt

Object 002

Series 1, Ask your store

Bulk update Shopify prices, see the list first

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: medium
Wow

Before

Repricing 40 variants by hand in the admin took an afternoon and two of them always came out wrong.

After

The table arrives in a minute, you approve it once, and the writes match the approved table exactly.

Fill in

[collection name]: the collection you are repricing, by its admin title
[percentage]: the increase you want, for example 8
[rounding rule]: how prices should land, for example the nearest .95

The one rule

Insist on the full before and after list before any write, or the agent will reprice archived and already discounted variants you never meant to include.

You should get back

A before and after table covering every variant in the collection, with archived and already discounted ones flagged in a status column. After you type approve, a second table confirms each write with the variant ID and the field written.

The prompt

Object 003

Series 1, Ask your store

Where is my order? Reply from the real order

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: easy
Wow

Before

Support replies were copied from a saved template that promised a delivery window nobody had checked.

After

Every reply carries the real fulfillment state, and a draft that would have promised a date the carrier cannot support gets flagged instead.

Fill in

[paste the customer email]: the full text of the customer's message, including the address they wrote from

The one rule

Make it quote the fulfillment fields it actually read, or you will send a delivery date the carrier never promised.

You should get back

The order's real fields listed with their names, then a short reply in a fenced block you can copy straight into your helpdesk. Anything the store does not hold, such as a carrier scan date, is named as missing rather than filled in.

The prompt

Object 004

Series 1, Ask your store

Analyse 200 customer reviews, find 5 complaints

Setup: No setup
Tools: Shopify connector for Claude, Reviews CSV export
Difficulty: easy
Wow

Before

The founder read one-star reviews as they arrived and formed an impression that changed every week.

After

Five themes come back ranked by count with quotes attached, so the scent-throw complaint stops being an argument and becomes a list.

Fill in

[number of reviews]: how many rows are in the export, for example 200
[attach the reviews CSV]: the export from your reviews app, with review ID, product, rating, date and body columns

The one rule

Set a minimum review count per theme, or you will rebuild a product around two complaints.

You should get back

Five ranked themes in a table with counts and date ranges, each followed by three verbatim quotes and their review IDs. A coverage line tells you how many rows were read and how many could not be classified.

The prompt

Object 005

Series 1, Ask your store

Translate a product page, keep every measurement

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: medium
Wow

Before

Marlowe and Vine had a machine translation of the candle range that turned burn time into something no customer could act on.

After

Translations come back beside the original with measurement lines preserved and the claims that need local wording flagged for a human.

Fill in

[product handle]: the product you are translating, by handle or exact title
[target language]: the language you are translating into
[target market]: the market the page is for, so claim rules can be flagged

The one rule

Pin every measurement and safety line to the original, or you will publish a page that is wrong in a way the English one is not.

You should get back

A side by side table of original and translation, field by field, with measurements carried across unchanged. Claims needing local wording are listed separately with a reason, and nothing is written to the product until you approve it.

The prompt

Object 006

Series 1, Ask your store

Which products have not sold in 180 days?

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: easy
Wow

Before

The catalogue had grown for three years and nobody had checked which items had quietly stopped moving.

After

One order-history pull flagged 123 never-sold SKUs, which turned a vague feeling about dead stock into a list with inventory value attached.

Real numbers from client work, names changed

Fill in

[none]: this one runs as written once the connector is installed; change 180 to a different window if your season is longer

The one rule

Make it state how it counted a sale, because refunded and cancelled orders can quietly keep dead stock off the list.

You should get back

A table of never-sold SKUs sorted by the money tied up in them, with last-order dates. Above it, the sale-counting rules in plain words, and below it a second short list of products excluded for missing data.

The prompt

Object 007

Series 1, Ask your store

Write alt text for 20 product images at once

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: easy
Wow

Before

Half the catalogue had alt text that repeated the product title and the rest had none at all.

After

Alt text describes the actual photo, screen reader users get a real description, and the table is reviewed before anything is saved.

Fill in

[product handles]: the handles of the products whose images need alt text, one per line

The one rule

Review the whole table before anything is applied, or you will save twenty near-identical strings built from the product title.

You should get back

A table of every image with its current and proposed alt text side by side, plus how each file was matched to a product. Images the agent could not see clearly are marked rather than guessed, and nothing is written until you confirm.

The prompt

Object 008

Series 1, Ask your store

Explain last week's numbers in plain English

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: easy
Wow

Before

The weekly check meant opening five dashboards and still not being sure which number had actually moved.

After

A short readout says what changed and what is noise, and every figure names the report it came from.

Fill in

[none]: run it as written every Monday; swap last week for a month if that is your reporting rhythm

The one rule

Ask which swings are normal week to week variation, or you will spend Monday chasing a slow Tuesday.

You should get back

Five or six plain sentences saying what moved, then a table of eight metrics with last week, the week before, the change and the Shopify report behind each. Swings the agent reads as ordinary variation are called out with its reasoning.

The prompt

Object 009

Series 1, Ask your store

Bulk edit 300 product titles to one format

Setup: No setup
Tools: Shopify connector for Claude, Shopify product CSV export
Difficulty: medium
Wow

Before

Three years of listings meant three naming habits, and onsite search returned the same candle three different ways.

After

One scheme, a file you can diff before import, and the ambiguous rows left untouched for a human decision.

Fill in

[row count]: how many product rows are in your export, for example 300
[attach the product CSV]: the product export from Shopify admin, all columns intact
[naming scheme]: the pattern you want, for example Brand, Product, Scent, Size

The one rule

Tell it to leave the handle column untouched, because a regenerated handle breaks every product URL you already rank for.

You should get back

Ten worked examples and the inferred rules first, then a CSV with only the Title column rewritten and the Handle column byte for byte identical. Rows too ambiguous to rename come back in a separate table with their original titles.

The prompt

Object 010

Series 1, Ask your store

Create a discount code, check the rules first

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: medium
Wow

Before

The last launch code stacked with an existing sitewide offer and sold a bundle below cost for two hours.

After

You see the configuration before it is written and the stored rules read back after, so the code is checked against what you asked for before the email goes out.

Fill in

[code]: the code customers will type, for example AUTUMN15
[percentage]: the discount value, for example 15
[collection name]: the collection the code applies to
[start and end]: the window, for example Friday 09:00 to Sunday 23:59

The one rule

Read the discount back from the store after it is created, or a default you never set, such as a combination rule, ships with it.

You should get back

The proposed configuration first, field by field, and nothing created until you approve it. After creation, a table of what the store actually stored beside what you asked for, with combination rules and other unset defaults called out.

The prompt

2 Connect the tools.

Someone wires Klaviyo, GA4, Search Console and your ad account in once, each through its own MCP connector. After that you ask one question that crosses two systems and get one answer.

Read the setup and all ten cards in Series 2

Setup 2

Series 0, Set up

Plug in the tools Shopify cannot see

Setup: Connector
Tools: Klaviyo MCP, GA4 MCP server, Search Console MCP server
Difficulty: medium
Wow

Before

Three dashboards, three logins, one spreadsheet.

After

One question that reads all three and shows where each figure came from.

Fill in

[nothing to fill in]: enable the connectors first, then paste this as written.

The one rule

Connect read-only or minimal-scope accounts first, and add write scopes only for the one card that needs them, when you get there.

You should get back

One row per connector with the account name and a proven number, or a clear failure message per connector.

Steps

Klaviyo: in Claude, add the Klaviyo connector and sign in with the account that owns the store's list.
GA4 and Search Console: use a connector that supports Google sign-in. Several vendors publish MCP servers, so pick one that lists the properties you own.
In each connector, choose the exact property or account for this store.
Run the prompt on this card.
Write down which connector is attached to which account. A card's fill-in will ask you for it later.

The prompt

Object 011

Series 2, Connect the tools

Audit my Klaviyo flows, starting with abandoned cart

Setup: Connector
Tools: Klaviyo MCP, Shopify connector for Claude, Claude
Difficulty: medium
Wow

Before

Marlowe and Vine had nine flows in the account, four switched off since a theme migration, and no abandoned checkout at all.

After

The inventory shows status and revenue per flow with the call behind each number, and the missing sequence arrives as a draft nobody activated by accident.

Fill in

[date range]: the window for flow revenue, for example the last 90 days
[flow to draft]: the missing sequence you want written first, for example abandoned checkout

The one rule

Say drafts only and no activation, or a flow can start emailing a segment nobody has looked at since last year.

You should get back

A table of every flow with status, trigger and attributed revenue, and a verdict column naming what is missing or switched off. Then three drafted messages in fenced blocks, created nowhere and scheduled for nothing.

The prompt

Object 012

Series 2, Connect the tools

Which pages lost money when rankings dropped?

Setup: Connector
Tools: Search Console and GA4 MCP server, Claude
Difficulty: hard
Wow

Before

The rank report and the revenue report lived in different tabs, so the fix list was ordered by position and the money was invisible.

After

The list comes back ordered by revenue lost, and the pages that could not be matched are named instead of silently dropped.

Fill in

[date range]: the recent window to compare, for example the last 90 days
[minimum sessions]: the floor below which you do not want a revenue verdict, for example 100

The one rule

Ask for the unmatched pages by name, or pages with tracking gaps drop out of the report and look healthy.

You should get back

One block showing the queries and property IDs, then a table of pages that lost both position and revenue, sorted by revenue lost. Pages present in only one source are listed by name, and low volume pages are marked rather than scored.

The prompt

Object 013

Series 2, Connect the tools

Transfer inventory between Shopify locations

Setup: Connector
Tools: Shopify Admin MCP, Claude
Difficulty: medium
Wow

Before

The second location kept selling out of the same three scents while the warehouse held months of cover.

After

A transfer table arrives with weeks of cover on both sides, and inventory only moves after a human has read it.

Fill in

[date range]: the sell-through window, for example the last 60 days
[low cover]: the point at which a location counts as running out, for example three weeks
[high cover]: the point at which a location counts as overstocked, for example eight weeks

The one rule

Forbid direct inventory adjustments, or you will spend a week reconciling stock that moved only on paper.

You should get back

A transfer table showing where each SKU would move, how much, and what weeks of cover both locations are left with. SKUs with no sales history or no inventory tracking are listed separately rather than folded into the plan.

The prompt

Object 014

Series 2, Connect the tools

Fill empty Shopify metafields so filters work

Setup: Connector
Tools: Shopify Admin MCP, Claude
Difficulty: medium
Wow

Before

Filters did not work because two thirds of the catalogue had empty metafields and nobody had time to fill hundreds of products.

After

Values come from the product's own copy, the unknowns are listed separately instead of guessed, and writing happens only after the CSV is approved.

Fill in

[catalogue scope]: which products to read, for example the whole catalogue or one collection
[metafield keys]: the metafields to fill, for example fit, fabric and size guide

The one rule

Require a source sentence for every value, or you will publish a fabric claim you cannot stand behind.

You should get back

A CSV of proposed values with the exact source sentence beside each one, and a second file listing every product where the copy said nothing. Writing happens only on the rows you send back, and each write reports its metafield ID.

The prompt

Object 015

Series 2, Connect the tools

Which tracking pixels are running on my store?

Setup: Connector
Tools: Shopify Admin MCP, Web fetch tool, Claude
Difficulty: medium
Wow

Before

Nobody could say how many analytics tools were live, only that page load had been getting slower every quarter.

After

The enumeration found three concurrent analytics stacks running at once, each with its own ID and its own version of the truth.

Real numbers from client work, names changed

Fill in

[read scopes]: the Admin API read scopes your MCP server was given, such as read_themes, read_script_tags and read_pixels; the default connector scopes do not cover these
[storefront URLs]: your home page URL and one product page URL, for the rendered check

The one rule

Check declared scripts against what the rendered page loads, or you will remove a tag that an app silently puts straight back.

You should get back

One table per source of truth, merged: every declared vendor and ID beside whether the rendered page actually requested it. Entries that could not be attributed to a file or extension are listed as unattributed rather than dropped.

The prompt

Object 016

Series 2, Connect the tools

Do my ad sales show up in Shopify orders?

Setup: Connector
Tools: Meta Ads MCP server, Shopify Admin MCP, Claude
Difficulty: hard
Wow

Before

The ads dashboard reported more purchases than the store recorded orders, and both numbers were used in the same meeting.

After

The two counts sit side by side with the attribution window stated, and the unexplained part of the gap is called unexplained instead of averaged away.

Fill in

[date range]: the window to reconcile, for example the last 30 days

The one rule

Fix the timezone on both sides before comparing, or the entire gap can be one day of orders landing on the wrong date.

You should get back

A side by side table of platform purchases against Shopify orders with the attribution setting and both timezones written out, then the gap broken into three sized buckets. The remainder is labelled unexplained with a number next to it.

The prompt

Object 017

Series 2, Connect the tools

Win back customers who stopped ordering

Setup: Connector
Tools: Klaviyo MCP, Shopify Admin MCP
Difficulty: medium
Wow

Before

The list held thousands of customers who bought once two years ago and had heard nothing since.

After

Six segments come back with counts and median order value, and two drafted sequences sit unsent until a human reads them.

Fill in

[recency bands]: the days-since-last-order buckets that match your buying cycle, for example 90 to 180, 181 to 365 and over 365

The one rule

Keep it to segment definitions and drafts, or a winback goes to customers who bought from you last week.

You should get back

Six cells in a table with counts, median order value and the products each band bought, plus the query behind the counts. Then three drafted messages for the two biggest cells, and segment definitions written for you to create by hand.

The prompt

Object 018

Series 2, Connect the tools

Check the product data Google and AI can read

Setup: Connector
Tools: Claude, Web fetch tool
Difficulty: medium
Wow

Before

The theme claimed rich result support, but nobody had opened the rendered markup since the last round of app installs.

After

Missing identifier and availability fields are listed per page with the raw markup shown, so the fix list can be checked line by line.

Fill in

[product page URLs]: five live product URLs, ideally from different templates
[collection URL]: one live collection page URL

The one rule

Read the rendered page rather than an app's claim, or you keep shipping product markup with no availability field.

You should get back

A per-URL table of schema types with missing required and recommended fields, plus any duplicate or conflicting blocks named. At least one page's raw JSON-LD is pasted back so you can check the reading against the source.

The prompt

Object 019

Series 2, Connect the tools

Know when to reorder stock, and how much

Setup: Connector
Tools: Shopify Admin MCP, Claude
Difficulty: hard
Wow

Before

Reorders were placed on the feeling that a scent had been selling well, and the best seller was out of stock for five weeks last winter.

After

A reorder list arrives with weeks of cover, the stated lead time and stockout days handled explicitly, ready for a human to price and send.

Fill in

[date range]: how far back to read sales, for example the last 120 days
[lead time]: how long your supplier takes, for example 45 days
[safety stock]: the buffer you keep, for example three weeks

The one rule

Exclude stockout days from velocity, or you will under-order exactly the SKUs that ran out.

You should get back

A reorder table with velocity, weeks of cover, excluded stockout days and a suggested quantity per SKU, followed by every assumption written out in sentences. A second table is a purchase order draft you can paste into an email and send yourself.

The prompt

Object 020

Series 2, Connect the tools

Track competitor prices every week, automatically

Setup: Connector
Tools: Web fetch tool, Scheduled run or Claude Code routine, Claude
Difficulty: medium
Wow

Before

Competitor prices were checked by hand before a launch and never again, so a rival's permanent price drop went unnoticed for a season.

After

A weekly diff lists only what moved, with the URL beside each change and failed fetches marked as failed rather than repeated.

Fill in

[competitor product URLs]: the product pages you want watched, one per line
[tracking sheet or file]: where the history is appended, for example a CSV in the project folder

The one rule

Record a failed fetch as not fetched and check robots.txt before scheduling, or last week's price gets carried forward as this week's fact.

You should get back

A short diff each week listing only the rows whose price or availability moved, with old and new values and the URL beside each. Pages that failed are recorded as not fetched with a status code, never carried forward.

The prompt

3 Give it hands.

Claude Code runs on your own machine with an Admin API key of its own: hundreds of edits in one run, done on a copy, shown to you before a single one reaches the store.

Read the setup and all ten cards in Series 3

Setup 3

Series 0, Set up

Give Claude Code its own key to the store

Setup: Claude Code
Tools: Claude Code, Shopify custom app, Admin API
Difficulty: hard
Wow

Before

Claude Code had no way into the store.

After

It has a scoped token, a staging theme, and a habit of stopping before it writes.

Fill in

[token variable]: the name of the environment variable holding your Admin API access token, for example SHOPIFY_ADMIN_TOKEN
[duplicate theme name]: a name you will recognise in the admin, such as playbook-staging

The one rule

Never paste the Admin API token into a chat window. Put it in an environment variable and rotate it the moment it leaks.

You should get back

The store domain, the scope list, two theme IDs, and nothing changed on the live theme.

Steps

In the Shopify admin, open Settings, then Apps and sales channels, then Develop apps, and create an app.
Configure Admin API scopes. Start with read_products, read_orders, read_themes and write_themes, and add write_products only when a Series 3 card needs it.
Install the app, reveal the Admin API access token once, and store it in an environment variable on the machine that runs Claude Code.
Install Claude Code and the Shopify CLI, then sign in to the store with the CLI.
Run the prompt on this card. It should end with two theme IDs and no change to the live theme.

The prompt

Object 021

Series 3, Give it hands

Make every product photo the same size and crop

Setup: Claude Code
Tools: Claude Code, Shopify Admin API, Image generation API
Difficulty: hard
Wow

Before

Product photos came from four shoots over three years, so the collection grid looked like four different stores.

After

108 packshots were recomposed to one framing and 104 went live after a three round review, with the four rejects never uploaded.

Real numbers from client work, names changed

Fill in

[packshot folder]: the local folder holding the original product photos

The one rule

Write to a review folder and never over the originals, or the one shot you cannot reshoot is gone before anyone reviews the batch.

You should get back

A contact sheet of the whole recomposed batch plus a per-file table showing the product it matched and any crop or colour issue. Nothing reaches the store until you send back approved filenames, and each upload reports its product and image ID.

The prompt

Object 022

Series 3, Give it hands

Fix blurry product photos across the catalogue

Setup: Claude Code
Tools: Claude Code, Shopify Admin API
Difficulty: hard
Wow

Before

Zooming a product image on a phone showed a soft, low resolution photo across most of the catalogue.

After

986 images across 505 products were upscaled and replaced in 20 chunks, each chunk verified on the live product before the next began.

Real numbers from client work, names changed

Fill in

[minimum long edge]: the resolution floor below which an image gets upscaled, for example 1600px

The one rule

Approve each chunk manifest and keep an undo manifest, or a mid-run failure leaves you unable to tell which products were already replaced.

You should get back

An inventory of every image below your resolution floor, then one approval prompt per chunk of about fifty. Each chunk is verified on the live product URL before the next starts, and an undo manifest maps every replacement back to its original.

The prompt

Object 023

Series 3, Give it hands

Build a local landing page for every city

Setup: Claude Code
Tools: Claude Code, Shopify Admin API
Difficulty: hard
Wow

Before

The store shipped to dozens of cities and had one generic delivery page that ranked for none of them.

After

53 localised city landing pages were created by script, each with its own copy and internal links, plus a rollback script that could unpublish exactly the pages it made.

Real numbers from client work, names changed

Fill in

[page template file]: the page template the copy is generated into
[cities CSV]: one row per city with delivery time, landmark and coordinate columns

The one rule

Demand a rollback script keyed to the handles it created, or a bad template cannot be cleanly undone at scale.

You should get back

A reviewable table of every page before anything is created, then the pages created unpublished with a rollback script keyed to those exact handles. Skipped rows are listed with the reason each was skipped.

The prompt

Object 024

Series 3, Give it hands

Edit your Shopify theme on a copy, never live

Setup: Claude Code
Tools: Claude Code, Shopify Admin API
Difficulty: hard
Wow

Before

The last theme tweak was made straight on the live theme on a Friday and took the cart drawer down with it.

After

Changes land on a duplicate with a diff and a preview link, and the live theme only ever changes when a human presses publish.

Fill in

[describe the change]: the theme change in one or two sentences, as specific as you can make it

The one rule

Work on a duplicate and publish it yourself, or your rollback is whatever you can remember on the day.

You should get back

A named duplicate theme, a full file diff, and a preview link. Alongside them, the agent's own check of what the diff could break in checkout, the cart or tracking, with the method for each check stated.

The prompt

Object 025

Series 3, Give it hands

Build Shopify retargeting ads from live stock

Setup: Claude Code
Tools: Claude Code, Meta Ads API, Shopify Admin API
Difficulty: hard
Wow

Before

Retargeting ads were built by hand from a spreadsheet exported weeks earlier, so sold out products kept collecting impressions.

After

The campaign was built from the live catalogue with 143 variants matched, 0 unmatched and 54 out-of-stock items auto-hidden, and it launched paused.

Real numbers from client work, names changed

Fill in

[product grouping]: how ad sets should be split, for example by collection, by price band or by category

The one rule

Drive product set membership from live availability rather than an export, or you will pay to advertise products nobody can buy.

You should get back

A full build plan as a table with match counts and the API calls behind them, reviewed before anything is created. Then a paused campaign with ad set and ad IDs and a preview link for each ad, plus a list of unmatched variants and why.

The prompt

Object 026

Series 3, Give it hands

Build a seasonal ad campaign, launch it paused

Setup: Claude Code
Tools: Claude Code, Meta Ads API
Difficulty: hard
Wow

Before

Every seasonal launch meant an evening in the ads manager and a typo in at least one budget field.

After

The campaign arrives built and paused with a preview for every ad, so the only manual step left is reading it and switching it on.

Fill in

[season or promotion]: what the campaign is for, for example an autumn range launch
[creative folder]: the local folder holding the images or video for the ads

The one rule

Require the full build plan before the first API call, or a mistyped budget spends real money while you are still reading.

You should get back

The build plan in full before any API call, including the optimisation event and why it was chosen. After approval, a paused campaign with four ad sets, every ID printed and a preview link per ad.

The prompt

Object 027

Series 3, Give it hands

Find the tracking pixel you never installed

Setup: Claude Code
Tools: Claude Code, Shopify Admin API
Difficulty: hard
Wow

Before

Nobody remembered installing half the tags, and a previous agency still had access to the theme.

After

The trace found a stale pixel that had been sending traffic data to an unrelated ad account for 9 months, and it was removed on a reviewed diff.

Real numbers from client work, names changed

Fill in

[accounts I own]: every analytics and ad account ID you actually control, one per line, so anything else stands out

The one rule

Trace each ID to an account you own before removing anything, or you will delete the one pixel your live campaigns depend on.

You should get back

An inventory of every tracking script with its destination ID and where it is declared, and anything reporting to an account you did not list flagged with a confidence note. Removals happen only on a duplicate theme, as a diff you read before it is written.

The prompt

Object 028

Series 3, Give it hands

Did the redesign actually work?

Setup: Claude Code
Tools: Claude Code, GA4 Data API
Difficulty: hard
Wow

Before

The redesign felt better and everyone wanted to call it a win from the first week of data.

After

With n=472 the test found one cut significant at p=.001 and the other flat, and the overall result was reported as inconclusive rather than as a win.

Real numbers from client work, names changed

Fill in

[redesigned page paths]: the page paths that changed, as GA4 reports them
[before window]: the dates before the redesign went live, matched in length to the window after

The one rule

Ask for sample size and the p value, or you will roll out a redesign on a week of noise.

You should get back

Sample size, conversion rate and revenue per session for both groups, then the test, the p value and the confidence interval with the GA4 query shown. When the result is not significant it is labelled inconclusive with the traffic needed to settle it.

The prompt

Object 029

Series 3, Give it hands

Score my Shopify store out of 100, rank the fixes

Setup: Claude Code
Tools: Claude Code, Shopify Admin API, GA4 Data API
Difficulty: hard
Wow

Before

Every consultant gave a different opinion on what to fix first and none of them showed their evidence.

After

The store graded 44/100 after four QA passes, every deduction carried evidence, and the items that could not be checked were listed as unchecked.

Real numbers from client work, names changed

Fill in

[store URL]: your public storefront URL, the one a shopper would land on

The one rule

Mark unchecked items as unchecked rather than failures, or the grade measures your own coverage gaps instead of the store.

You should get back

The rubric first, for approval, then a score with a section breakdown and a findings table where every deduction carries a URL, screenshot or query result. Unchecked items sit outside the denominator, and the fix list is ranked by money with unsizable items marked.

The prompt

Object 030

Series 3, Give it hands

Fix signups that never reach your email list

Setup: Claude Code
Tools: Claude Code, Shopify Admin API, Customer data platform API
Difficulty: hard
Wow

Before

The popup looked like it was working, but the signups that appeared in the store never arrived in the customer platform.

After

The ordering bug was patched on a duplicate theme, 249 customers were backfilled, and a 12x under-report was corrected.

Real numbers from client work, names changed

Fill in

[page URL]: the page where the signup popup appears
[start date]: how far back the backfill should reach
[match key]: the field used to match a Shopify customer to a downstream profile, for example email

The one rule

Run a dry run before any backfill, or a wrong identifier writes hundreds of duplicate profiles you then merge by hand.

You should get back

The observed network calls in order with their payloads, so the ordering bug is visible rather than described, plus a theme diff waiting for your approval. Then a dry run of the backfill, and per-batch counts from both systems as they converge.

The prompt

4 What went wrong.

These cost real money. Four stores, four bad afternoons. Names changed, numbers kept.

Read all four cards in Series 4

Object 031

Series 4, What went wrong

Traffic but no sales: 17,722 sessions, zero orders

Setup: Connector
Tools: Meta Ads MCP server, Shopify Admin MCP
Difficulty: medium
Wow

Before

The ad set reported healthy add-to-cart volume at a good cost per action, so nobody looked further down the funnel.

After

17,722 sessions and 0 completed orders in 30 days, caused by an ad set optimising for add to cart in a market with no working payment rail.

Real numbers from client work, names changed

Fill in

[date range]: the window to check, for example the last 30 days

The one rule

Compare the optimisation event against completed orders, or the platform will keep buying you people who cannot pay.

You should get back

One table per ad set showing spend, the optimisation event count and real Shopify orders beside each other. Ad sets firing the event with no orders behind it are flagged, with the payment method and shipping zone for each targeted country checked and named.

The prompt

Object 032

Series 4, What went wrong

Meta ads say sales, Shopify says none

Setup: Connector
Tools: Meta Ads MCP server, Shopify Admin MCP
Difficulty: medium
Wow

Before

Two reports, two purchase counts, and a monthly meeting that used whichever one supported the point being made.

After

Every order is traced to the systems that recorded it, and the part of the gap nobody can explain is labelled unexplained.

Fill in

[date range]: the window to reconcile, matched on both sides

The one rule

Reconcile platform purchases to real orders before scaling spend, or you will scale against sales that did not happen.

You should get back

Three order ID lists showing what each system saw, then a summary table sizing duplicates, window effects and orders the platform missed. The remainder carries the word unexplained and a number, not a split difference.

The prompt

Object 033

Series 4, What went wrong

The AI made up a number, and it looked right

Setup: No setup
Tools: Shopify connector for Claude
Difficulty: easy
Wow

Before

The assistant answered instantly with a conversion rate that sounded plausible and matched nothing in the admin.

After

Numbers now arrive with the report or query attached, and the questions that cannot be answered come back as could not retrieve.

Fill in

[your question]: the thing you actually want to know, asked once the source rule is set, for example what was my conversion rate last week and the week before

The one rule

Require a source before any number, or you will run a quarter on a benchmark the model remembered from somewhere else.

You should get back

A table where every number carries the report, date range and timezone it came from, and anything the connector could not return sits in the same table marked unavailable. No benchmark and no estimate appears anywhere in the answer.

The prompt

Object 034

Series 4, What went wrong

A bulk edit with no preview, and no undo

Setup: Claude Code
Tools: Claude Code, Shopify Admin API
Difficulty: medium
Wow

Before

One instruction, one confident agent, and a catalogue-wide update that finished in a few seconds.

After

The write stops at a list you approve, runs in verified batches, and leaves an undo file built from the values it replaced.

Fill in

[the change you are about to make]: the bulk edit in one line, for example set the vendor field across 400 products

The one rule

Require a full preview list and an undo file, or the fastest thing an agent does is the thing you cannot take back.

You should get back

A complete preview list, not a sample, with the current and proposed value for every record, then a hard stop for your approval. Writing runs in verified batches of fifty and finishes with an undo file and the command to run it.

The prompt

Questions.

What is a Shopify AI agent?

A Shopify AI agent is an AI assistant connected to your live Shopify store, rather than one you paste screenshots into. Through Shopify's official connector it reads your real orders, products, inventory and customers, and with write access it can change them. The model matters less than the connection. On every job here, a person still approves the writes.

How do I connect AI to Shopify?

Install Shopify's official connector from inside ChatGPT or Claude, then approve an access level in your Shopify admin. That takes about five minutes and needs no developer and no token. Add an MCP connector only when the assistant has to read a tool Shopify does not own, such as Klaviyo. Use Claude Code when the job writes hundreds of records.

What is in the Shopify Agent Playbook?

37 cards: 30 jobs a connected assistant can do, 3 setups that unlock them, and 4 things that went wrong on real stores. Every card names the tools, what the prompt is allowed to change, the one rule to keep, and what you should get back. The prompts and the PDF are free with an email.

How much does the Shopify Agent Playbook cost?

Nothing. Every card's summary, tools and guardrail are on the page for anyone to read, and one email unlocks the prompt text plus the PDF of all 37 cards. There is no paid version and nothing to install. If you would rather have the work done than read it, the Shopify store scan is free too.

Is it safe to let an AI agent change my Shopify store?

It is safe when the order is fixed: read first, then the full list of proposed changes, then your approval, then batches you can undo. Every card states its mode, so you know before you type whether the prompt only reads, drafts something unsent, writes a CSV you import yourself, or writes with your approval.

Is there an official Shopify MCP server for Claude Code?

No single product covers it. Shopify ships a Dev MCP server for documentation and API schemas, a Storefront MCP for shoppers' agents, and the official connector for chat assistants. None of those gives Claude Code admin access from your own machine, which is why Setup 3 creates a custom app carrying only the scopes Series 3 needs.

How do I get started with Shopify MCP?

You may not need it. If you want to run the store from a chat window, the official Shopify connector is the whole answer, and it needs no token and no developer. MCP earns its place when one question spans two systems, such as Klaviyo flow revenue next to the products that actually sold.

How do I get started with a Shopify MCP server?

That one is a developer job, and the delay is never the code. Before anyone starts, decide which Admin API scopes the work actually needs, where the server runs, and who owns the token and rotates it. Keep write scopes off until a read-only version has answered the question you bought it for.

Which AI assistants work with a Shopify store?

Four, for different jobs. Shopify Sidekick works inside the admin and comes with your plan. ChatGPT and Claude use Shopify's official connector, so they also reach your files and your other tools. Claude Code runs on your own machine and is the one that does bulk work. Shopify publishes a connector for Perplexity as well.

What is agentic commerce on Shopify?

Two directions, often confused. Outward: shoppers' own AI assistants can read your catalogue and build a cart, which comes with being on Shopify rather than being something you switch on, and Shopify calls it early. Inward: you point an assistant at your own admin to answer questions and make changes. The second is the 30 jobs here.

Want it done, not read?

Run the free store scan and we will show you where the money leaks first.

Run the free scan