The problem
Vendors send product lists with missing specs, inconsistent names and the occasional wrong value. Filling the gaps by reading datasheets is slow, and a language model that researches fast will also confidently fill a gap with a guess. On a live store, a wrong specification costs more than a missing one: a customer buys the wrong part, or a listing is overwritten.
Who needed it: the catalog team, and anyone who later has to trust a product page.
How it works
- Ask I name the products in chat and say whether to build, fix or look something up.
- Research Researcher agents find the manufacturer's document for each part and save it with the date it was fetched.
- Write Page copy, specifications and attributes are drafted from those documents, under a 78-line rulebook of house rules.
- Check Scripts check the format and that every number in the output appears in the source it cites.
- Verify A second model looks only at the cells the scripts flagged.
- Images One clean, square image per distinct picture, shared across a product family so the store's library stays tidy.
- Hand over A WooCommerce import file, the sources used for each part, and a short note with questions for me. I review it and upload it.
This is the second generation. The first took a vendor's entire spreadsheet through four agent roles and a verification gate: 5,025 cells, 1,909 of them proven against a source, 978 blanks filled and 48 vendor errors corrected. It taught me what the evidence standard had to be, and the lean tool above is the result.
Decisions that made it fast and safe
- Scripts check first; models only look at what scripts flag. The original design had a model verify every row. Reordering the work cut the cost of a build by two thirds with no loss in accuracy.
- Telling the model to be frugal made it cost more. I ran five variants against a golden set with cached sources and a measured noise floor, and that instruction went. A cheaper writer model saved less than the run-to-run noise, so that went too.
- Blank beats a note. No "Unknown", no "not specified". A cell holds a sourced value or nothing, and anything uncertain comes to me as a question instead.
- A fix touches only what it was asked to. Fix mode hands the agent a slice and merges back only those columns, so every other cell in the file stays byte-for-byte identical.
- The store is unreachable by construction. Permissions deny any write to the live store, and a tripwire test proves it. Upload is always done by a person.
- It refuses to guess a match. A prefix, suffix or near match to the catalog is a candidate for me to confirm, never an exact match.
- Hostile sources do not steer it. A source page planted with "ignore the rulebook" instructions was ignored in every test run.
The first production batch
In late September the tool built 174 products in one batch. Of 2,455 values it wrote, 2,363 were confirmed word for word in a cited manufacturer document, and the other 92 came to me flagged with the reason.
Along the way it caught things a person skimming would miss. One vendor listed a pack of 10 as a pack of 50, and the manufacturer's document said otherwise.
It also caught a spreadsheet lookup error that would have shipped as a store ID.
Results
My part
I direct coding agents to write the code. I write the plan, the constraints and the stopping conditions, design the experiments, write the rulebook from what I know about the products, and rule on every conflict the tool raises.
Ongoing: bringing existing listings up to the same standard.
- Stack
- Python, pytest, coding agents and skills, WooCommerce CSV import
- Repository
- Private; inputs are company data.