McKee Outdoors: a website two years in the making, finished in six days

McKee Outdoors is an independent fishing tackle shop in Maryville, Tennessee. For over two years the owner had tried to load his inventory onto his website one item at a time. From Sep 28 to Oct 3, 2026, I directed AI agents that loaded, researched, connected and cleaned his catalog, and left him an AI assistant to run the store from his phone.

Fishing tackle cut from grape and grey paper on black card: two lures, a spinnerbait, three hooks, a spool of line and an open tackle tray. — Cut-paper illustration made for this page, not a photo of the shop.

Summary

Role: Independent project, Sep 28 – Oct 3, 2026. I ran it solo, working with Claude through Claude Code. I set the goals and rules, approved each risky step, and handled everything that needed the owner.
Client: McKee Outdoors, an independent fishing tackle shop in Maryville, Tennessee, owned by Matt McKee. A Lightspeed Retail (X-Series) register and a Shopify website.
Results: Live products went from 513 to 2,483, in 23,260 sizes and colors. 91% of them take price and stock from the register automatically. 7,637 barcodes that would not scan were repaired and 4,389 duplicate register items merged. The store is complete and ready to launch.
Tools: Claude through Claude Code · Lightspeed Retail (X-Series) API and import files · Shopify Admin API and CLI · the Lightspeed–Shopify connector · an MCP server on Cloudflare Workers for the owner's assistant

Problem

McKee Outdoors stocks over 12,000 different items. When the work began, its website showed 513 of them, and 480 of those had no photo. The menu had three items.
Nothing kept the website's prices and stock in step with the register. The register's built-in connector only synced the listings it had created itself.
Entering a product by hand takes 10 to 20 minutes with photos and options. For the 6,849 in-stock items missing from the website, that is months of full-time work, and it never catches up with new stock. The website stayed unfinished for over two years.
The register's own data was thin and messy. Only 4% of in-stock items had a description. About 10,400 barcodes had lost their leading zero, so packages did not scan, and thousands of items existed twice.

What I did

I ran the project solo, working with Claude through Claude Code. The AI studied both systems' APIs, wrote and ran about 90 data scripts, and sent out about 250 research-agent runs. I set the goals and the rules, approved each risky step, and handled everything that needed the owner.
Six ground rules held for the whole project. The register is the source of truth for price and stock. Back up before every change. Test on one item, then a small batch, then run in bulk. Archive, never delete. The owner decides. Never invent a fact; where the AI was not sure, the item went on the owner's checklist instead.
I loaded a distributor's catalog, 3,555 products, as hidden drafts. The categories were rebuilt into 134 collections with a three-level menu, maker photos were added, and the theme's leftover demo content was removed.
Before the Lightspeed–Shopify connector touched the live store, nine controlled single-product tests proved how it behaved: what it overwrote, when it silently dropped products, and what triggered a price push. Products were then linked in batches of 100 or fewer, each batch verified on four counts: linked, price matches, stock matches, web address unchanged.
AI research agents took the 6,849 missing items in order of sales. Each agent took a whole product line, found the exact item from its barcode on the maker's own website, matched each register item to a color or size, picked the maker's photo, chose a category and wrote an original description. A separate checker confirmed every item, and any eight-word overlap with the maker's text was rewritten.
Connecting the two systems exposed problems at the register that cost money at the counter every day. Each one was measured, planned with the owner, tested on a small batch and then fixed in bulk, mostly after hours.
For the hand-off, the project built a 63-tool AI assistant the owner uses from the Claude app on his phone, on a private MCP server on Cloudflare Workers. Every change previews first and waits for a yes. Six plain-language guides and a tick-off checklist went with it.

What changed

The website went from 513 live products to 2,483, in 23,260 sizes and colors. The menu went from three items to 134 collections.
2,399 of the 2,483 live products, 91% of the sizes and colors, take price and stock from the register within minutes. The final batches verified at 100% on all four counts.
1,229 brand-new products were researched, written, given the maker's photos, built on both systems and linked. 2,636 colors and sizes the shop carried but the website did not show were added to existing products.
At the register, 7,637 barcodes that would not scan were repaired, 4,389 duplicate items were merged with their stock combined, and 252 items priced at $0 or below cost were repriced to the maker's suggested price.
The owner runs the store by talking to his phone. The website is complete and waits behind a maintenance page until he works through a short checklist and switches it on.

Before and after

Measure	Before, Sep 28	After, Oct 3
Live products on the website	513	2,483
Sizes and colors for sale	1,664	23,260
Products without a photo	480 of 513	Nearly none
Website categories	A three-item menu	134 collections, three-level menu
Website synced to the register	Only the original listings	91% of sizes and colors, within minutes
Barcodes that would not scan	About 10,400	7,637 fixed; the rest were duplicates, now merged
Duplicate register items	4,418 pairs	29 pairs left for a shelf count
Items at $0 or below cost	305	52 left, each with a reason
Owner tools	None	A 63-tool AI assistant, six guides and a checklist

How the six days worked

Back up everything, then match the website, the register and a distributor catalog item by item
Load the catalog as hidden drafts and link it to the register in verified batches
AI research agents fill the gaps from the makers' own sites, with an independent check
Fix the register's data: barcodes, duplicates and prices, after hours, with the owner's OK
Hand over an AI assistant on the owner's phone, six guides and a checklist

The storyboard has five stages. The first day backed up every product, photo, collection and setting, and matched the website, the register and a distributor catalog item by item. The distributor catalog was loaded as hidden drafts, and products were linked to the register in small verified batches once nine tests had proved how the connector behaved. AI research agents then filled the gaps from the makers' own websites, with a separate checker on every item. Next the register's own data was fixed: barcodes, duplicates and prices, mostly after hours and each with the owner's approval. The last stage handed the owner an AI assistant on his phone, six guides and a checklist.

Figures come from the project's own logs and live system data, checked as the work ran. The shop's sales and inventory values are left out.

Related case study: Service Impact Detection. An AI network analysis tool that went into production.

If your shop has the same problem, or you're hiring for this kind of work, send a message. The full case study is available on request.