skill
last-job-search
An AI agent system on Claude Code that runs your whole job search: finds and scores roles, tailors and fact-checks your resume, finds who to message on LinkedIn, and reminds you to follow up.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/aranculleton/last-job-search itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 25, 2026last-job-search
The last job search tool you'll ever need. Claude finds jobs that fit you, scores each one, writes and checks a resume and cover letter for the ones you choose, finds the people worth talking to and drafts what to say, keeps track of every application and conversation, reminds you when to follow up, and gets you ready for interviews. It all happens on one private page that Claude keeps up to date: you tap, and Claude does the work. It never sends anything. You stay in charge of what goes out.
Find jobs → score each one → you choose → write and check → find the people → follow up
What you need
- A GitHub account (free).
- Claude Code, which comes with a paid Claude plan. On your own computer an Anthropic API
key works instead and you pay only for what you use, often cheaper if you search now and
then, but the scheduled brief and cloud chats need a plan.
- Another agent tool (ChatGPT Codex, Google's Gemini CLI, the Cursor editor)? The core works in any agent that can read files and run commands; AGENTS.md covers the Claude Code parts.
- No coding, and nothing to install first. A phone is enough.
Get started
- Make your own private copy: on this page, Use this template, then Create a new repository, with Private chosen. Never fork it: a fork is public.
- Open claude.ai/code, connect GitHub when it asks, and install the Claude GitHub App on your copy only.
- Choose your copy and say "set me up". Claude asks a few plain questions, does the research and opens your page with your first shortlist on it.
START_HERE.md has each step for a phone, a browser or a computer, and HELP.md what to do if something goes wrong.
What it looks like
What you do on your page
Tap, or say it in any chat. Claude does the work in the background and shows its progress.
| You do | You get |
|---|---|
| Search again | New jobs, each scored 0 to 100 against you, with the reasons |
| Draft on a job you like | A resume and cover letter written for that job, checked |
| Add a job (paste a link or the posting) | The same, for a job you found yourself |
| Prepare on an interview | Likely questions, your best answers and questions to ask |
| Contacts | People worth talking to at the employers you want, and a first note for each. You send it |
| To do | Replies and follow-ups that are due, drafted for you |
| Tweak on anything Claude made | A change of look, length or tone, before you approve it |
| "They replied" or "I got an interview" | The record updated, and what to do next |
No resume yet? Say "build one with me". It never asks for your email: new jobs and what is due reach you on your page and in your twice-weekly brief.
Why it works differently
- It ranks, it does not hide. A job advert is a wish list. A missing skill lowers a score; it does not throw the job away. Only a few things rule a job out, such as needing a right to work you do not have, or a language you do not speak. Rulings based on your rules are re-checked when the rules change; ones in your own words stay until you change your mind.
- It works your network, not just the job boards. It looks where people in your field are actually named (team pages, meetups, speaker lists, alumni), says in a line why each person is worth writing to, and writes every message in the shape that suits them: the first note (which asks for a conversation, not a job), the replies and the referral ask. It tracks every thread and warns you before a cold application lands on top of a warm introduction.
- It reminds you. When a follow-up is due, on an application or a conversation, it says so and drafts the nudge. Setup ends by making your brief: what you owe, what changed and the next three things to do, in one message, twice a week.
- It checks its own work. Every resume is compiled, read back and checked the way an applicant tracking system reads it, and every claim against what you told it.
- It remembers. Applications, conversations and what is due live in files in your copy, not in a chat that ends.
- It learns what you like. Say "too formal" or "stop suggesting recruiters" once, and it sticks. It also notices which messages got replies and which applications got interviews.
- It keeps itself up to date. Each release reaches your copy by itself the next time you use it, even in a chat kept open for weeks. Nothing about you is ever sent back.
Your data and security
Everything about you stays in your private copy, and nothing is submitted for you: every document stops as a file, and you send it. Job postings are written by strangers, so their text is treated as data, never as instructions, and a link inside one is never opened. A check fails any change that weakens the permission allowlist or lets personal files into the app. SECURITY.md has the details.
For the technically curious
You never need any of this to use it.
| Set up a computer or a cloud environment by hand | SETUP.md |
| What a search does, step by step | RUNBOOK.md |
| How each stage works and what it costs | ARCHITECTURE.md |
| The rules every chat follows, and where every other rule lives | CLAUDE.md |
| The twice-weekly brief and scheduled searches | automation/README.md |
| Scoring rubric, writing style, cost model | .claude/skills/job-application-assistant/ |
| Reporting a bug or suggesting something | CONTRIBUTING.md |
| Working on the app itself | MAINTAINING.md |
Searching, filtering and scoring are plain Python and Bash, so a search reads thousands of
postings without using the model. The model starts only when you pick what to write. The
tests run in CI on macOS and Linux, and on any copy through ./tools/ci_local.sh. The
employer-board reader is also published as openroles.
Credits and licence
Built on MadsLorentzen/ai-job-search (v1.7.0 plus the commits through PR #413), which supplied the LaTeX scaffolding, the job-board CLI pattern, the skills-and-commands layout and the drafter-reviewer idea. Everything before the shortlist, the bullet allocator, the checks and the networking layer were built here.
MIT. LICENSE carries both copyrights. Not affiliated with Anthropic.