Before Claude Code writes pages, clean product data

A previous piece treated the product master — one row per product — as the single source of truth, with Claude Code writing separate pages for each mall. This time the focus is the first stumbling block: notation drift. A sample table for a fictional store shows the same mug entered three ways, with "ホワイト", "白" and "WHITE" for color and "450ml", "４５０ｍｌ" and "250mL" for volume.

These differences have three practical effects: shoppers searching one term may miss products written another way, byte-counted character limits can be exceeded by full-width entries, and AI output tends to mirror the inconsistent input. The proposed fix is to insert a dedicated cleanup step between the product master and the mall-specific writing step — not to fold it into generation — so the corrections stay visible.

The method adds rules to CLAUDE.md, writes to products_clean.csv without overwriting the original, and outputs a list of every changed item. Style codes, JAN codes, brand names and product proper nouns are excluded from automatic correction and sent to a "要確認" list instead. EC自動化工房 says a "商品ページ量産キット" bundling the templates and rules is scheduled for release on 10月31日, with the announcement to be shared on X at @ec_jidoka_kobo.