Steve Ward

CS 270

Sample-Scan Agent

An agent that scans product campaigns, pulls per-item diligence data, and delivers a ranked shortlist to my phone, then stops. The design point is the brakes, not the speed.

For
Solo operator running a product-review business
Format
Agent skill: browser research pipeline with hard action limits

The problem

Product research for an influencer business meant hours of manual campaign scanning: checking identifiers, prices, ratings, and monthly sales for every candidate, over and over, before any asking or outreach could happen.

What I built, by directing AI

An agent skill that scans a product category, pulls per-item diligence data, and delivers a ranked shortlist to a phone. Then it stops. The interesting engineering is the boundary, not the scan.

Guardrails

The agent can look but never act. It is forbidden from accepting offers, messaging brands, or exposing credentials or addresses. An exclusions file controls what it skips. Its first runs happened only with the operator watching, in an isolated browser session that gets closed when the work is done.

How I measured it

The output is a ranked shortlist with source data per item, which a human acts on. The operator's decisions become ground truth for judging future ranking quality. Baseline: two hand-run scans delivered five-item ranked shortlists with diligence data attached.

What happened

Installed and enabled, verified on real scans. Scheduled autonomy is deliberately withheld until trial runs prove reliability.

What I'd do next

Compare the agent's rankings against realized outcomes over a few cycles before letting it run unattended.

Back to the catalog