The Claude operating system for Shopify merchants

The official Shopify Claude connector landed this week and it is the upgrade operators have been waiting on. The connector reads orders, products, inventory, customers, and discounts directly. The first time you watch it close 30 minutes of admin work in one approval tap, you will get it.

The connector is the moment everything got faster. The architecture in this post is what makes it compound.

Most operators hit the same wall around month two. The morning chat that worked in week one starts mixing decisions across tasks. The voice doc lives in four different versions across four chats. Memory is off because nobody turned it on. Connector or no connector, that wall is real.

The fix is the system underneath the prompts. Get it right and the connector stops being a faster way to ask the same questions, and starts being the operator stack you have been building toward.

Claude has 5 context layers: profile preferences, project instructions, knowledge base, skills, memory. Each one holds a different kind of context, and getting the assignment right is what makes the whole system compound.

Profile preferences is account-wide. Brand voice, hard nos, role context. Loads automatically into every conversation in every project.

Project instructions is per workstream. Identity for that folder, usually one or two sentences telling Claude what role to play inside the project.

Knowledge base is per project. Reference data Claude needs to do the work. SKU dictionaries, campaign archives, KPI definitions specific to that workstream. With the connector on, the knowledge base shifts from holding raw data to holding the rules and definitions that turn raw data into decisions. Lighter, sharper, more useful.

Skills are reusable workflows triggered by slash command. Built once, runs from any project. With connectors on, skills graduate from "ask me for a CSV" to "pull the data and act."

Memory is passive synthesis that compounds. The longer it runs, the sharper it gets.

Most of these patterns come from putting context in the wrong layer. Brand voice in a project instead of profile preferences. KPI definitions duplicated across project knowledge bases. Skills written from scratch every time instead of saved as reusable.

Get the assignment right and the system stops drifting. With the connector on, the architecture pays back even faster, because every prompt now reaches live data through clean rules.

Universal context goes in profile preferences. Project-specific reference data stays inside that project.

Voice rules apply everywhere → profile preferences.

KPI dictionary applies everywhere → profile preferences.

SKU dictionary applies only to inventory work → Inventory project knowledge base.

Campaign archive applies only to campaign work → Campaigns project knowledge base.

Never duplicate. If something belongs in profile preferences, it doesn't go in any project. If something belongs in one project, it doesn't get copy-pasted into another.

Bloated knowledge bases cause Claude to ignore the rules underneath. Shorter and sharper beats long and complete. With connectors on, the rule is even sharper, because the connector pulls the live data, the knowledge base only needs to hold the rules.

When everything goes into one chat, inventory questions, campaign drafts, customer replies, by week three Claude starts mixing decisions across tasks because the context window is full of unrelated work.

The fix is structural rather than prompt-level. Each chat works better as a working session for one job. Done, close, next. The exception is iterative drafts where you're building one piece across multiple sessions, and those stay open until the piece ships.

Naming chats after the task makes them searchable two weeks later when you're trying to find that brief you wrote. "Draft Q4 launch brief" beats "campaigns" every time. Generic names become noise.

Brand voice gets re-pasted into every new conversation, and three months in there are usually four versions of the same voice doc floating around in different chats with none of them matching. Each one was written when context was fresh and the voice felt obvious, but they drift apart over time.

Profile preferences is account-wide, which means whatever you write there loads automatically into every conversation in every project. The setting lives in Claude's account preferences and is easy to walk past on signup.

Four sections do most of the work.

Operating context is who you are, what brand, what scale, what team you work with. This is the basic setup that every conversation should start with.

Voice rules is sentence case, no em dashes, fragment style, whatever your brand demands. The clearer the rules, the less Claude drifts.

Hard nos is banned phrases, banned structures, never invent numbers, never write press-release copy. Negative rules are often more effective than positive ones because they tell Claude what not to do.

Pushback bias is how aggressive you want Claude to be in disagreeing with you. Default Claude is agreeable, and at operator scale agreeable is expensive.

A short profile preferences doc beats a long bloated one. After a month of running Claude with it, you'll notice patterns that need adjustment. Tighten then. Cut anything that isn't doing work.

Most operators leave memory off and the common reasons are privacy concerns, not sure it's relevant, or didn't realize it was there.

The cost is that every Monday start cold.

With memory running 90 days, Claude builds a working synthesis of how the business runs. Seasonal patterns, top SKUs, repeat customer behavior, the agencies in rotation, the KPIs that matter and the ones that don't.

Leave it on, audit it monthly.

Ten minutes is all it takes. Open settings, view what Claude remembers, edit out stale frames.

The audit checklist is short. Delete old goals you've moved past, old org structure if a key person left, closed workstreams Claude is still referencing. Keep voice rules that are working, long-term context still true, patterns Claude picked up across hundreds of conversations.

A folder named "Marketing" with random files dropped in is a topic, not a workstream. Too broad to load with useful knowledge, and too generic for Claude to have a working identity inside.

Workstreams are streams of work that repeat on a schedule. Daily ops repeats every day. Performance review repeats every Monday. Campaigns repeat every month. Each one earns its own folder with its own knowledge and its own instructions.

These cover most ecom operators regardless of role.

Daily ops holds yesterday's sales, ad spend, email performance, inventory exceptions, and the prior-week decision log. The folder you open every weekday at the same time.

Performance review holds the last 12 weeks of revenue, ad ROAS by channel, email engagement, agency scorecards, and the KPI dictionary. Pre-call prep for weekly meetings.

Campaigns and calendar holds monthly plans, creative library, copy that landed and didn't, the holiday calendar.

Inventory and merch holds SKU sell-through, abandoned cart by SKU, collection performance, slow-mover history, margin bands.

Customer holds loyalty rules, top spenders, CX macros, voice of customer themes, win-back history.

Personal IP holds top 30 LinkedIn posts that landed, voice rules, past keynotes, internal memos. Anything that goes out under your name.

Agency management if you run external partners. Competitive intel if you have the bandwidth to track competitors. Capital and finance if you're a founder or COO managing the P&L.

Nine projects max. More than that, folders start pretending to be tags.

Skills are reusable workflows you build once and trigger from any project with a slash command, and the same prompt fires every time but the output gets sharper as memory builds underneath.

/morning-brief reads yesterday's data and gives you a decision queue in 12 minutes; three things to act on, one thing to worry about.

/weekly-prep surfaces patterns across the last 4 weeks (not just this week vs last week) and drafts the questions to ask each function lead on the call.

/sku-alert flags inventory that crossed a threshold. Slow-movers that need a discount decision, fast-movers running out of margin to reorder, abandoned cart spikes that need a marketing pivot.

/campaign-postmortem pulls spend, revenue, AOV, CAC vs LTV impact for any campaign and compares it to the brief.

/reply-drafter takes a customer ticket, classifies it, pulls the right macro, adapts to tone, and outputs three voice-clean variants.

/eod-log captures today's decisions, what changed, and what opens for tomorrow. Memory feeds on this, and tomorrow's brief reads sharper because of it.

Once the core six are stable (usually around day 30), add the advanced ones: cohort report, supplier review, returns analysis, agency scorecard, board prep, decision retro.

This is the part that just got better.

Shopify shipped the official Claude connector. The old setup, downloading Claude Desktop, running terminal commands, editing CLAUDE.md by hand, is no longer the starting point. Settings, Connectors, Shopify, install. It installs an app inside your store and you are connected in under a minute.

The connector exposes 27 tools across 4 permission tiers, and the design is the part that makes it usable on a real store from day one.

Read paths auto-approve. Get Product, Get Order, List Customers, Run Analytics Query. No prompts, just answers. Ask "what's my abandoned cart trend this week" and Claude pulls the numbers in line.

Interactive tools fire on intent and pause to show the plan. Create Product, Update Product, Set Inventory, Create Discount. The plan shows up first, you tap to confirm, the change lands. The approval step is what makes it feel like having a senior ops person on staff who runs everything by you before they touch the store.

Write/delete tools confirm every call. Raw GraphQL mutations and Switch Shop. The wider the action, the cleaner the confirmation flow.

App-only tools sit in their own bucket for the new store generation flow.

The connector is fast and it does exactly what you tell it. The catalog underneath is the leverage point. Fill the metafields, name the variants well, and every connector prompt gets sharper from there.

Walkthrough video below.

Brand voice doc lives in Drive. Update it once, every project pulls the latest version.

Shopify connector reads orders, products, collections, metafields. Operators can ask "what's my abandoned cart trend this week" and Claude pulls the actual data instead of asking for a CSV.

Gmail connector drafts replies in your tone, summarizes long threads, searches your inbox. Important: it drafts only. Won't send on your behalf.

Calendar connector pulls this week's meetings into the morning brief. /weekly-prep gets sharper because Claude knows what calls are coming up.

Connectors inherit your account permissions. If you can't access a file, the connector can't either.

If you'd rather skip the setup and just have it running, I take on a small number of build engagements per quarter. Two-week sprint, built around your team and your catalog, including a Loom walkthrough so your team can run it without me afterward.

Reach out at [email protected] if it sounds useful.

Your catalog is the foundation of every project in this stack.

Atomz runs three products that automatically handle the AI visibility layer. Different jobs, same goal.

This tells you what AI agents see when they crawl your store, and where you're missing.

Run your free audit here

Most catalogs aren't structured for AI agents. Missing attributes, vague descriptions, no context for material, fit, use case. Catalog Genius enriches every product with the attributes ChatGPT and Perplexity actually match against. Runs automatically. No manual SKU work.

Schedule a demo to fix your Catalog

On-site search that works the way shoppers ask. Natural language queries, semantic matching, attribute-aware results. Runs alongside your existing Shopify search or replaces it.

Fix your on-site search with Atomz AI search

Do you run something like this and want to compare notes?

I spend most of my week deep in agent orchestration, model routing decisions, and what's actually working on Shopify catalogs at scale. If you're building in this space, I'd rather trade notes than give a talk.

LinkedIn: https://linkedin.com/in/ankitminocha

Email: [email protected]

Direct links to the settings and docs referenced in this post.

claude.ai/settings - Click your initials in the lower-left corner, go to Profile, find "What preferences should Claude consider in responses?" Available on every plan.

Anthropic's memory docs - How chat search and memory work, including how to pause or reset.

Claude Code skills documentation - Full reference for SKILL.md files, frontmatter, and slash command invocation.

Awesome Claude Code on GitHub - Community-curated skills, hooks, agents, and slash commands worth borrowing from.

Compare Claude plans - Pro, Max, Team, Enterprise. Compare context window sizes and connector availability.

Anthropic Enterprise overview - Connectors, org-wide projects, the 500K context window.

Connectors Directory - Browse all available connectors including Shopify, Google Drive, Gmail, Calendar, Notion, and Microsoft 365.

Anthropic's connectors documentation - How to enable connectors, set action restrictions, and manage permissions.

Google Workspace connectors guide - Specific setup for Gmail, Calendar, and Drive including read-only permissions.

Written by Ankit Minocha, founder at Atomz AI (https://atomz.ai)