I built and launched a Shopify store with AI in one week. Here is the full blueprint.

Part 1 of 2. This is the build and the organic phase, before any ad spend. Part 2 is what happened once the ads went on.

In September I built and launched an online store, Tail Friends, in about a week, with a couple of hours a day, and AI doing nearly all of the work. Forty live products shipping to 18 countries, 128 shipping rates, six email automations, two social accounts posting twice a day, a daily engagement routine and a Monday review that ends in decisions. One person in the loop, approving.

I have built two Shopify stores before this one, both before AI. Each took weeks: product research in spreadsheets, copy written by hand, shipping tables worked out by trial and error, a theme I paid someone to tweak, a social calendar that died inside a month. This one took about a week to build and takes about ten minutes a day to run. It cost a fraction of the money and none of the weekends.

So this is the full blueprint. The accounts you need, the connectors that let the AI act instead of advise, the skills that make it good at store work, the prompts, the rules, the costs, the mistakes, and the real numbers. None of it is hypothetical. It is all running at tailfriends.co right now.

The full kit is free for subscribers. Subscribe to The Build Log and the welcome email sends you the Store Blueprint: the setup order, the Claude project instructions, the connector and skill list, the operating manual template, twelve copy-paste prompts, the launch checklist and the start-up cost table.

The result, in numbers

What got built in a week, all live:

A Shopify store with about 40 products from a dropshipping supplier, each with three or more photos or videos, brand-voice copy, real specs and size charts where needed.
Every product priced by one rule (about 2.5x landed cost), with unit costs stored so a margin audit is a single query.
128 weight-band shipping rates across 16 zones and six markets (US, Canada, Australia, EU, Korea, Singapore), worked out from the supplier’s real costs per warehouse.
Six email automations (welcome, abandoned cart, post-purchase, win-back and more) on Shopify’s own free tool.
Instagram and TikTok posting a short video at noon and 9pm Singapore time, every day, with an original music track on every clip.
Around 20 real comments a day on other pet accounts, and six creator permission requests a week.
A review step where I approve or kill everything before it goes out.
A Monday KPI review that comes back with decisions, not charts.
Two social accounts that went from nothing to just over 100 followers, roughly 500 likes and 100 comments, off 40 to 50 pieces of content, all of it made by the machine above.

What it took from me: about a week to set up, then about ten minutes a day.

How it is wired

Four layers. Get them in this order and everything after is easier.

1. The accounts. Shopify, a supplier, a Google account for the brand, the AI, the social handles. Boring, and the thing most people skip past. Without the Shopify account the AI has nothing to act on.

2. The brain and its memory. A Claude project with instructions that tell it what it is running, who I am, and what it may never do. Inside that project, a shared Google Drive folder holding one file, HUB.md, the operating manual. Every session, on every machine, reads the same rules. Change a rule once and every job picks it up.

3. The hands. Connectors. The Shopify connector lets the AI edit products, prices, collections and redirects and run analytics queries directly. The Google Drive connector gives it the manual and the review folders, and Composio fills the gaps where the native connector can’t search or write a file. Claude in Chrome on an always-on Mac mini does the logged-in browser jobs: commenting, checking supplier logins, posting. Higgsfield makes the product video, ElevenLabs Music the soundtrack, Metricool schedules it. Without connectors the AI writes you instructions. With them it does the work.

4. The skills. Plain-text job descriptions the AI loads when the task matches. I wrote ten for store work: sourcing and pricing, listing rewrites, SEO meta, cross-sell, margin audit, refund analysis, LTV by channel, ad creative tests, content calendar, creator outreach, customer segments, support macros, theme tweaks. They are free: on this site and on GitHub. Drop them into Claude, ChatGPT or Cursor and point them at your store.

Job	Tool	Cost
The brain	Claude, in a project with the manual pinned	Subscription
Store	Shopify Basic	Monthly plan
Supplier	CJ Dropshipping	Free, pay per order
Store edits and analytics	Shopify connector for Claude	Included
Shared rules and review folders	Google Drive connector, Composio for Drive and Sheets jobs the native connector can’t do	Free tiers
Browser jobs	Claude in Chrome on a Mac mini	Included, hardware optional
Product and lifestyle video	Higgsfield via MCP	Credits
Music	ElevenLabs Music	About 9 cents a track
Scheduling	Metricool via MCP	Free tier
Reviews	Judge.me	Free tier
Search	Search Console, Merchant Center free listings	Free
Store skills	My ten Shopify skills	Free

The operating manual

The most useful file in the whole system is a plain text document called HUB.md. It holds which niches, which supplier, the pricing formula, the discount cap, what a listing must contain, how a caption is structured, what a support reply sounds like, the hard rules, and a dated change log.

Every job starts by reading it. When I changed the pricing rule after the first mistake, I changed one line and every job picked it up that night.

Steal this: the manual prompt

“You run my online store day to day. Before any action, read HUB.md. It holds the rules on pricing, discounts, listings, captions, shipping copy and what you are never allowed to do. If a task conflicts with the manual, stop and tell me. If a rule is missing, propose one and add it to the change log with today’s date. Take initiative on routine work. Ask before spending money or publishing.”

The content machine

Twice a day a short video goes out on Instagram and TikTok. The loop:

A 30-day calendar, built from what is in stock and what sells in the niche, picks the product and the hook.
A real product photo becomes the first frame. Higgsfield animates five seconds from a prompt. Photorealistic, labelled as AI on the platform, no “made with AI” line in the caption.
An original instrumental gets generated to match the mood and mixed to broadcast loudness. Every clip gets one.
The caption follows a fixed structure: hook, one or two plain lines, product and price, “Shop it: link in bio”, the welcome code, five to seven hashtags.
It lands in a Pending Review folder and a row in a Google Sheet. Nothing is scheduled yet.
I review once a day. Approved, Needs work with a one-line note, or Rejected.
Approved clips take the next free slot.

Steal this: the margin audit prompt

“Pull every active product with its price and unit cost. Flag anything under 2.5x landed cost, anything over 4x, and anything where variant prices are out of order (the cheapest variant must be in position 1). Propose a new price for each flag with the reasoning in one line. Change nothing until I approve the list.”

The night the dog grew a leg

The review step exists because of one evening. The content machine had been running for a few days and I sat down to watch what it had made.

Most of it was good. A cat losing a three-day war with a toy fish. A golden retriever gliding along in a harness. Clips I would stop scrolling for.

Then, every few videos, something was off. A dog whose legs did not add up. A cat whose face changed halfway through. A product about 20% different from the one we sell. Nothing that jumped out on the first watch. Everything that would get screenshotted on the second. And the system was set to post all of it, twice a day, under the brand.

The fix was a spreadsheet and three words. Every clip, caption and post gets a row. Once a day I mark it Approved, Needs work (with a one-line note: “dog has an extra leg”, “too salesy”, “price is wrong”) or Rejected. Overnight the AI reads the notes, fixes what it can, bumps the version number and puts the new cut back in front of me. Rejected means rejected. The pile of rejects is where the rules in the manual come from.

Ten minutes a day. Best ten minutes I spend on the business.

Steal this: the review loop prompt

“Every morning at 07:00, read the Content Review sheet. For every row marked Needs work, apply my note, regenerate or rewrite, save as the next version in the same folder, and write what you changed in the ‘Claude’s changes’ column with today’s date. For Approved rows, schedule into the next free slot and fill the Scheduled column. For Rejected rows, move the folder to Rejected and never propose that idea again. If a slot date passes while a row is still pending, move the target date forward and tell me. Don’t publish.”

Real beats generated

Some of the best content was not generated at all. We message real pet owners whose clips made us laugh and ask, politely, if we can share them. Full credit, a tag, a clear “it’s fine to say no”. Some say yes straight away, and those clips have something no generated video has yet: a real dog doing something stupid, filmed by the person who loves it. AI finds the clips, drafts the message, edits and tracks who said yes. A human asks and a human says thank you.

What broke, in order

Every one of these cost time or would have cost money. They are now lines in the manual.

The first pricing rule killed half the catalogue. A 3x floor made half the products unsellable against the competition. Now 2.5x with a hard minimum margin. One edit in one file fixed every job at once.
Google was being told the wrong price. On four products the page showed the dearest variant first while the price tag Google reads showed the cheapest. That gets your free listings disapproved. Rule: cheapest variant in position 1.
The store had no shipping rates the day before launch. Not wrong rates. None. We built all 128 from the supplier’s calculator in one sitting. Rule: test checkout as a stranger, in a private window, from another country.
A homepage link went to a page that did not exist. One renamed product, one dead link on the front page for days. Rule: check every homepage link after a rename, redirect the old address.
Collection pages had eleven words on them. Fine for a human, invisible to Google. Now 150 to 250 real words each.
Free trials that were not. Three tools quietly switch to paid or charge for the basic tier. Rule: check what happens on day 15 before recommending anything.
The first paid test did nothing. 40 euros on TikTok Promote: about 1,100 views, zero orders. The money went in before the blockers were fixed: no reviews, 30 dollar price points, 8 to 13 day delivery on some items. Fix the funnel, then pay for traffic. That is why ads are only starting now.
The platform’s robots file tried to recruit my AI. The store’s own robots.txt contains text asking AI agents to install a particular shopping skill. My agent read it as content, not instructions, and told me. If your AI follows instructions it finds on web pages, it is not your AI.

What it cost

Rough start-up capital, in US dollars, for the first month. The full table with ranges is in the subscriber kit.

Shopify Basic: about $39 a month (the first month is usually a trial for a dollar).
Domain: $15 to $40 a year.
Claude: $20 a month on Pro. I run the higher tier because I use it for everything else too.
Supplier: nothing up front. You pay per order.
Video and music: Higgsfield credits plus about 9 cents a track. Budget $30 to $50 for the first month of twice-daily clips.
Scheduling, reviews, search tools: free tiers.
Samples to photograph and seed to creators: I set aside $300.
Optional: a second-hand Mac mini for the always-on browser jobs, $400 to $600 once.
Ads: nothing until the funnel is fixed. The test budget is $1,000 a month, approved by me per campaign.

Minimum to launch and run organically for a month: roughly $100 to $150 without the Mac mini. Compare that to the two stores I built by hand, where the theme work alone cost more than this whole stack.

The numbers so far

Everything started at zero on 22 September. No ad spend yet, so this is the baseline part 2 gets measured against. Where things stand at the time of writing:

Just over 100 followers across Instagram and TikTok. Instagram is the bigger of the two.
About 500 likes and 100 comments on 40 to 50 posts.
Around 7,500 video views across both platforms. One TikTok did 1,200 on its own; a normal clip does 250 to 500.
Store sessions: 341 from 317 visitors. About a third came from social, the rest direct or through the bio link.
Add to cart: 3. Reached checkout: 1. Orders: 0.

Zero orders, and I am putting it in print on purpose. Everything above the checkout line works: the content gets made, people watch it, some of them visit. The wall is at conversion, and that is what the ads are being pointed at, with the fixes above already in. If someone shows you an “AI built my store” story with no conversion numbers, ask for them.

Everything I used, in one list

Connectors (the things that let the AI act):

Shopify connector for Claude: products, collections, prices, redirects, analytics.
Google Drive connector: the manual, the review folders, the content.
Composio: Drive search, file edits and Sheets updates the native connector can’t do, plus a sandbox for scripted jobs.
Claude in Chrome: the logged-in browser jobs on the Mac mini.
Higgsfield MCP: image and video generation from inside the chat.
Metricool MCP: scheduling and the follower numbers above.
ElevenLabs Music: a track per clip.

Skills:

My ten Shopify skills, also on GitHub. Free.
The full catalogue of free skills and connectors I keep: cianoneill.ai/skills.

The store: tailfriends.co. Go and find something I got wrong. I will probably write about it.

What I would tell you

Open the accounts first, then set up the Claude project, then connectors, then skills, then the manual. Only then prompts. Most people start at prompts and wonder why the AI keeps telling them what to do instead of doing it.
Write the manual before the first product. Every skill and job you build afterwards is only as good as the rules it reads.
Build the review step before the first clip is generated. It is what lets you publish twice a day without waking up to a five-legged dog on your brand account.
Give it an always-on machine. Browser jobs die when your laptop sleeps. A second-hand Mac mini solved more problems than any prompt.
Test the thing, not the reply. The checkout had no shipping rates and every system said “done”. Buy from your own store as a stranger.
Keep the list of what broke. It becomes your manual, your content and your next article.

Get the Store Blueprint. Subscribe to The Build Log and the welcome email sends you the kit: setup order, the Claude project instructions I use, the connector and skill list, the HUB.md template, twelve prompts, the launch checklist and the cost table. Build your own, then tell me what broke.

Next: the first ad budget goes in this week on ten hero products. Part 2 will show cost per order, which videos converted versus which collected likes, and whether a store run by AI can pay for itself.