Jobs machine
- Solo: product design, brand and the full build
- 5 days, 15–19 November 2026
- Next.js, Supabase, Vercel, Claude or Gemini
Fewer applications, better ones. An AI job search tool I designed and built for my own search. It reads every posting, scores it against my real work, rules out the jobs I can't legally take, and drafts the application. Then it stops. I read it, edit it, and press send myself.
The problem
Job hunting had become a reading job. Dozens of postings a day, most of which turned out, on paragraph six, to be US-only, a junior role in disguise, or tied to a time zone I couldn't work in.
The AI tools I tried made it worse. They were built for volume: auto-apply to hundreds of jobs, stuff the CV with keywords, send it in your name. That gets you filtered out, and it puts claims in your CV you can't defend in an interview.
The bet
Let the machine do the reading and the drafting. Keep the judgement and the sending with people.
Nothing is sent without you
No auto-apply, no emails in your name. Applying on the employer's own page is always yours.
Only your real facts
The writer may only use what is in your profile. Every claim in a CV traces back to a case study you wrote.
Your key, your data
You bring your own AI key. No credits, no recruiters.
Decisions
- Eligibility is not part of the score. Can I take it and do I want it are different questions. A separate chip rules a job out before you read it; the score only measures fit.
- Score against case studies, not titles. Designers are judged on work, so setup asks for case studies in words.
- Skip asks for a reason. One tap ("salary", "too junior", "agency") turns each no into a record of why.
- Marking the machine's voice. Anything the model wrote starts with ✳. Text quoted from a posting never does, so you always know who said what.
The system
- Gather. Open job boards, Y Combinator companies, Hacker News threads, RSS feeds and careers pages on Greenhouse, Lever and Ashby.
- Judge. A fast model reads each posting into facts and scores it 0–100 against your profile, with reasons and gaps.
- Prepare. A stronger model drafts a CV, a cover letter and form answers. Each part can be redone with a one-line correction.
- You apply. You read, edit, and apply on the employer's own page, then follow it on the Tracker.
Highlights
- Rework
Three layouts, one chosen
Split view, a dense ledger and a focus mode, built as clickable HTML. Split view kept both the list and the "why".
- Fit readout
A number on a square tile
A red, amber, green bar read like a school grade. Fit is now one tile with four sub-scores under it.
- Teams
Share the jobs, not the CVs
A team shares the job pool and what they saved; CVs, letters and notes stay private.
- Identity
A match, not a jackpot
Three reels with the middle one landed dead centre. Six rounds, generated from a small script.
Looking back
The hardest and most valuable call was the feature I refused to build.
Not auto-applying shaped the positioning, the trust model and the interface. I used the first version for real before redesigning it, and that produced a better brief than any persona would have.
The full case study covers the rework, the visual system and the identity.
Next in the lab
Choose, Don't Generate
A practice tool for design decisions: pick between directions, then defend the choice.
View the case