When you connect Claude to your Shopify store and ask it to update 50 product titles, it will do it. All 50 - done immediately. Without showing you what it changed, without asking if you meant all 50 or just the ones in a specific collection. Later, when something looks wrong, you have no way to trace what happened.
The CLAUDE.md setup I am going to share fixes this. It is a folder structure of instruction files that Claude loads at the start of every session. It makes Claude reliable and efficient.
It works for a 20-product store as well as for a 10,000-product store. The structure is the same and all you need to do is fill in the details.
1. The folder structure
Everything starts with one file: CLAUDE.md (attached below).
This is the entry point Claude reads at the start of every session. It defines what the brand is, what role Claude is operating in, and which other files to load. The rest of the setup only works if this file points to it. Inside the .claude folder, there are four subfolders.
Rules - how Claude reads data, how it writes, and what requires approval before acting
Commands - pre-saved prompts you can call by name instead of typing the full thing each session (things you’d want to do/check regularly)
Hooks - actions that fire automatically after certain operations (logging, for example)
Agents - for multi-step operations. Not covered in this session.
The structure is simple which is intentional - the goal is something you can actually maintain, not something you configure once and forget.
I have recorded a video which you can watch instead - follows the same order.
2. Read rules
Before any write can go wrong, you lose tokens. Read rules define the defaults for how Claude retrieves data from your store. This sets up a small, useful result set which can be limited to 10 or 20 with key fields only and sorted by most recent.
The reason this matters: Claude will return everything it can access if you do not tell it not to. On a large store, that is a lot of tokens consumed before you get a useful answer.
Read rules also change by role. Someone managing the catalog has different defaults than someone pulling customer analytics. The rules file is personal to how you operate, not universal - and you can use this as a starting point to build on top of it.
3. Write rules
Writes are where mistakes become real. The write rules enforce this sequence before anything changes in the store: confirm scope → show what would change → run on a small set first → show the result → then run on the full scope
Three specifics worth noting.
Tags: always fetch the current tags on a product before writing new ones. If you do not, Claude may replace the existing tags with the one you just asked it to add.
Metafield dictionary: keep a separate file listing every metafield your store uses, the namespace, and what it does. Without this - Claude writes to whichever field seems closest, and with it, it writes to the right one.
Error handling: if a mutation fails mid-operation, stop immediately. Report the count of what succeeded, what failed, and the IDs of the failed items.
4. Approval gates
This is the layer that protects the store. Four tiers in the file I am sharing:
Reads and dry runs: no approval needed
Single item writes: one quick confirm and only then proceed
2 to 10 items: show a dry run first and only proceed when you say yes
11 or more items: always stop and ask, no exceptions
And three rules that override everything regardless of scope:
Any price change stops. It does not matter what else is happening in the operation. If a price update is triggered anywhere, Claude stops and surfaces it.
Any URL change stops as it affects SEO. They are not catalog edits.
Any deletion stops
In a session with Claude, the approval gate caught a scope mismatch before touching a single product. The request was for 15 updates and Claude saw 21 active products, flagged the contradiction and waited for clarification. That is how the gate works.
5. Commands
Once the rules are in place, commands are how you use the setup day to day.
A command is a long prompt saved under a short name. Instead of typing the full audit prompt every morning, you type daily audit - Claude runs the whole thing.
The setup in the video includes five example commands.
daily audit - inventory health, orders, metafield status, top three things that need attention today. Claude surfaces the priorities without being asked.
fill missing metafields - finds any products where a specific field is empty and fills it using the brand voice file and store context.
pull report - a custom analytics view formatted for your workflow.
safe bulk update - a bulk operation with approval gates already built into the prompt.
You can build as many as you need and for different teams can have different commands. The commands replace the repetitive part of any operation with Claude.
6. Brand voice and policies
Two context files that load at the start of every session alongside the main CLAUDE.md.
The brand voice file is where your actual brand guidelines live. What words belong in a product title, what never goes in a description - the tone, style, and what the brand is not. Think of it as training Claude on how your catalog should read.
The policies file holds the URLs for your return policy, terms and conditions, refund policy and anything customer-facing. Small use case: if someone on your team asks Claude to draft a reply to a customer complaint, it now has the right policy URL to reference.
Neither of these files does anything on its own - they work because CLAUDE.md tells Claude to load them before doing anything related to them.
The file
The full folder used in the video is attached below. You need to:
→ Replace the brand placeholders with your store
→ Adjust the approval gate thresholds to match your catalog size
→ Add your metafield dictionary - the structure is the same across stores
Before you set any of this up - if you haven't checked how visible your store is to AI shopping agents, run a free audit at gpt.atomz.ai. Takes 30 seconds. It'll show you exactly what ChatGPT sees when someone searches for your products.
Ankit M.