Nate DouglasSenior Product Designer, home

The tool behind my job search.

After I got laid off, I spent a little over 45 days searching on my own. I applied to over 300 jobs, about 9 every weekday, and didn’t hear back once. So I designed and directed a job search pipeline that screens postings with plain rules, tailors my resume to each match, and hands every account wall to a person.

Role
Designer and director. Claude Code wrote the code.
Built with
Claude Code, Python, SQLite
Timeline
Since mid-October 2026, still running
34drunning since October 12, 1:54 PM ET
  1. 368applications sentabout 14 every weekday
    300+ resumes tailored
  2. 41recruiter touches22 replied to an application
    11 reached out first
    8 virtual screening invites
  3. 18screens or interviewsfrom recruiter touches
    virtual invites never count
How these were counted

From the pipeline’s own database, October 12 to November 16, 2026. Resumes tailored is a floor, because packages for rejected and skipped jobs are deleted. Recruiter touches add up every company whose recruiter contacted me, counted once whether or not we talked, plus virtual screening invites. 22 recruiters replied to applications the pipeline prepared and 11 reached out on their own; 18 of those became screens or interviews. Virtual screenings are automated invites that most applicants get, so they count as touches but never as screens. These are my own job search numbers, not users of a product.

Senior Product Designer, Payments

Contoso Ledger, remote

88%keyword match
floor 75%

Seven years on payment and scheduling flows, including an invoice redesign tested with users. The main risk is no A/B testing.

Not added
A/B testing. It isn’t true, so it shows as a gap.
Verdict
Clean. The package agrees with the master resume.

Resume PDFPostingRe-tailorI appliedSkip

What the review page shows for each match, with Sam Example from the public repo.

How I worked

Setup

  1. Rules firstWrote my own search rules for title, pay, place and posting age as plain filters, so every rejection has a reason I can read.
  2. Base resumesApproved three base resumes, so tailoring edits a few bullets instead of rewriting the whole thing.
  3. Claude Code, headlessPrompted Claude Code with every tool turned off, and treated posting text as untrusted.

Guardrails

  1. Truth checksEvery number, employer and date gets checked against my master resume before I see it.
  2. The gateThe last check before Submit fails closed, and a person handles every account wall.
  3. Failures become rulesEach bad form or posting became a rule, and in the public repo, a test.
Problem

Applying early matters. Tailoring by hand doesn’t keep up.

  1. Good roles fill fast, and a tailored resume took me up to 20 to 30 minutes.
  2. Generic AI tailoring invents things: new numbers, and titles I never held.
  3. Application forms carry stale autofill, account walls and questions only I can answer.
The idea
Claude doesn’t grade its own work.

The model writes. Plain code checks. The keyword score is computed locally, every number has to already exist in my master resume, and anything the checks can’t confirm comes to me.

Honest tailoring

Small edits by bullet, never new facts.

Claude edits bullets by ID on a base resume I already approved. The headline stays a title I actually held, no employer can be dropped, and a keyword only goes in if it’s true. This example uses Sam Example.

Master resume

Redesigned the invoice payment flow for 40,000 small businesses; usability test task success rose from 61% to 88%.

A draft that tried to improve it

…task success rose from 61% to 94%.

Blocked: 94 isn’t in the master resume.

A keyword the posting wanted

A/B testing

Not added, because it isn’t true. It shows up on the review page as a gap instead.

The gate

Four forms, four verdicts.

The gate reads a snapshot of the filled form. Any reason at all means a person looks, so a wrong “no” costs a minute and a wrong “yes” doesn’t happen. These are the real verdicts on the demo forms in the repo.

Clean form

Clean

A person can press Submit.

Account wall

Needs you

The site wants a new password. Creating accounts is never automated.

Stale autofill

Stop

First name says Jamie, should be Sam. An old resume is still attached.

Sponsorship

Stop

The form says the candidate needs visa sponsorship. They don’t.

What I cut
The public repo stops at the decision.

The version I run also fills forms in a browser. Across many applicant tracking systems that was the most fragile part, and anything that signs in or sends email touches my own accounts. So the public code answers one question: given what the form shows, can it go?

What I’d do differently

Draw the human handoff first.

Start with one file of facts

My first version spread personal facts across rules, prompts and settings. One file, read everywhere, is where I’d start.

Write the failing case first

Most rules began as a posting or form that got through when it shouldn’t have. I’d write the example first, then the fix, so every rule has a test.

Measure outcomes sooner

The keyword score is a proxy. I added reply tracking a week in, and I’d build it before tuning the score at all.

Define the gate before automating

The hardest cases were account walls and odd custom questions. I’d decide where a person steps in first, then automate only what passes cleanly.