Inspired visual search skill.

Inspired visual search skill

An agent skill that teaches a coding or chat agent to shop a photograph with Inspired: choose the right focus domain, start one durable read, poll it without burning the user's allowance, and report what came back in words the pipeline can actually support.

One file, SKILL.md, with YAML frontmatter. No scripts, no network access of its own, nothing that runs on installation.

Why it exists

The three mistakes an agent makes with this API are predictable:

1. It leaves the domain on fashion and asks a living-room photo about cushion covers.
2. It spends the allowance. Starting a read costs one request from a ten-request trial key, so a retry loop, a duplicate call for the same image, or a "get a fresh key" workaround empties it in a minute. Polling is free and agents often do not know that.
3. It overstates the result. Nothing in Inspired verifies a cutout against the original photo or scores how close a retailer match is, so "I found your sofa, it is this one" is a claim the product cannot back. The skill fixes the vocabulary: objects are spotted, results are visually similar.

What it covers

• A table mapping what the photo shows to the focus domain to send, and when to call list_focus_domains instead of guessing.
• What focus is for — narrowing within a domain — and why it cannot be used to redefine the reader's instruction.
• Starting a read over MCP or over POST /api/v1/jobs, with the image constraints and why POST /api/v1/inspire is never the answer.
• Polling: honouring pollAfterMs, and how much to say to the user while a multi-minute read is in flight.
• A worked output format, and the required vocabulary.
• The awkward cases handled individually: an empty result, the wrong objects, a spent allowance, a 404 on a collect, a failed read.

Who it helps

Anyone running an agent that will be handed photographs — a shopping assistant, a catalogue enrichment loop, a research agent working through a moodboard.

Limits

It teaches the platform as deployed at inspired.jetskibay.com in September 2026. It assumes a key already exists in the environment or on the MCP connection and never tries to mint one. It does not make an agent better at visual search in general and it grants no access on its own.

Here is a photo of my living room. I'm redoing it and I want to know what the
lamp and the sofa would cost me — find me somewhere to buy each, or something
close enough.

[attach: living-room.jpg]

I have an Inspired key set up on the MCP connection. Two things before you
start: don't run the search twice, I only have ten of them; and be straight with
me about how close the results actually are, I don't want to order a sofa
because a photo looked similar.