subscribe.txt× Close

Get the next desktop in your inbox.

A new desktop every week.

No spam. Unsubscribe anytime.

Want your desktop here or know someone who should be featured?

Leave your email and we will reach out.

You are looking at one person's real AI setup. A new desktop every week.Subscribe

What is Korab building?

Between roles, building in public while the job search runs

Used-car search kit

Visit github.com

Screenshot of Used-car search kit

A kit for buying a used car where the AI does the research and I make the decisions. It maps nearby dealers, finds matching cars, checks their histories and drafts questions for the seller, then stops for my OK at each step.

Why
Agents are fast at the tedious parts of a car search, but left alone they fill gaps with guesses, overstate progress and act too early. I wanted the speed without a car I hadn't checked showing up as 82 out of 100.
How long
Published October 2026

How I built it

  1. I wrote a one-page search brief in plain English: where I live, how far I'll drive, my budget, the models and my must-haves.
  2. I wrote rules the agent reads first: never guess a missing detail, show how much of each car has actually been checked, never contact a seller.
  3. I split the search into gates, so the agent stops for my approval before each step that costs money or commits me.
  4. I added a small local research service with a scoring rubric I can change.
  5. I scrubbed my real search and kept it as a worked example.
To copy it you need
Claude Code, Codex, or the AI of your choice that can read files and browse the web. Copy the template on GitHub and fill in the search brief. Gotcha: tell the agent that unknown means unknown, never 'no' and never zero.

Try it yourself

Copy this prompt and paste it into ChatGPT or Claude to try it yourself.

I'm buying a used car and want my AI to do the research while I make every decision. Korab Eland built a kit for this, shown on whatarewebuilding.ai (github.com/korabeland/used-car-search). Korab writes a plain-English search brief, and the agent maps nearby dealers, finds matching cars, checks histories and drafts questions for sellers, stopping for approval at every step. Its key rule: an unknown stays unknown. It never becomes 'probably' or a zero, and scores show how much was actually checked. Help me do the same. Start by asking me the brief questions one at a time: where I live, how far I'll drive, budget, models and must-haves. Never contact a seller for me. Ask me for anything you need.

Tools I used for this project

Next file: Perian

Claude CodeCodexGeminiOllamaHermes AgentObsidianSupabaseVercelNotionFirstmate