Product Safety Scan Skill — Development skill for Claude Code

Product safety scan — a Claude skill

Hand Claude photographs of a product label. It reads the ingredient panel and both company names, checks the ingredients against a list **you** wrote in advance, searches the manufacturer and the importer separately for recalls and regulatory action, tests whether both businesses actually exist and ship things, explains any warning on the package in plain language, and returns 🟢 / 🟡 / 🔴 with the single finding that produced it.

It tells you what is findable and what it means. It does not certify that anything is safe, and it will say so if you ask it to.

What it checks

• The label — ingredients against a list you write, allergens, warnings.
• The two companies — manufacturer and importer are usually different firms in different countries with different records. Searched separately.
• Recalls across jurisdictions — FSANZ, UK FSA, CFIA, Hong Kong CFS and EU RASFF as well as the FDA, for the whole brand family rather than the exact SKU, with a check on whether recalled stock is still inside its best-before dates.
• The ground it grew in — for spices, herbs, tea, rice and other crops that concentrate soil metals, the growing region rather than the head office, and whether anyone has published a measurement of the commodity itself.
• The warning, against every threshold that applies — a Prop 65 exposure trigger, an EU concentration ceiling, the national limit where it grew, and any published risk assessment. These disagree by design, and the disagreement is the finding.

The verdict rule

**Worst finding wins.** One confirmed recall sets red. One company you cannot verify sets red. Nothing averages it back down, however clean the rest of the scan came out.

Rating systems drift toward the mean because averaging feels fair. Averaging is the wrong operation for hazard, where the question is never how a product scores overall but whether anything about it is disqualifying.

`avoid-list.