The Hottest Job-Search App Is a Repo You Fork

When Mads Lorentzen's geophysics position was cut in late 2025, he did what a certain kind of engineer does: he built tooling for the problem. Sixty-nine tailored applications and twenty first interviews later, he signed a contract as an AI engineer, then published the whole pipeline as ai-job-search under MIT. It hit #1 on GitHub trending in July and now sits at roughly 33,000 stars with nearly 12,000 forks.

The job-search framing is what got it trending. It's not the interesting part. The interesting part is that one of the fastest-growing repos of 2026 isn't a SaaS product, a web app, or even really a program. It's a folder of markdown instructions, slash commands, and small CLI tools that only does anything when you drop it into Claude Code. The tagline — "fork it and own it" — is a distribution model, and I think it's the one a lot of personal software is about to adopt.

What you're actually forking

Strip away the career-coach framing and the architecture is a legitimately well-built agent pipeline. /setup interviews you and builds your profile into CLAUDE.md and supporting skill files — plain markdown, version-controlled, greppable, no database. /scrape runs job-portal searches through small Bun CLIs that live under .agents/skills/, each following a standard output contract, then deduplicates and ranks results by fit. /apply <url> is the centerpiece: it evaluates the posting against your profile, drafts a CV and cover letter in LaTeX, then spawns a second Claude agent with fresh context to research the company and critique the drafts before the first one revises.

That drafter–reviewer split is the same pattern production agent teams have converged on — a critic that hasn't seen the drafting context catches things the drafter rationalizes away. So is the verification obsession. Compiled PDFs get visually inspected and iterated until the layout is clean, and the CV's embedded text layer is extracted with pdftotext and checked the way an applicant tracking system would parse it. When a CV overflows two pages, lines are scored by relevance to the posting and cut lowest-first rather than oldest-first. Lorentzen is explicit that this burns extra tokens in exchange for fewer broken documents, which is the correct trade and one most agent demos skip.

Notably, there's no auto-submit and no mass-apply mode. That distinguishes it from the spray-and-pray bots — the LazyApply-style browser automators and the auto-apply agents that trended in 2024 — which optimized for application volume and helped teach recruiters to distrust anything AI-shaped. This tool optimizes quality per application and leaves the send button to you. Lorentzen's own funnel is refreshingly honest about what that buys: 69 applications to 20 interviews is a strong hit rate, and 20 interviews to one offer is a reminder that no pipeline interviews for you.

The runtime is your Claude subscription

Here's the structural bit worth sitting with. This app has no backend, no accounts, no API keys, and no telemetry. The runtime is Claude Code itself — the agent CLI a large share of its target audience already pays for. Your CV, salary expectations, and interview post-mortems stay in a local git repo. Extensibility is /add-portal, which generates a new job-board integration matching the skill contract, and community forks maintain portals for their local markets.

Two years ago "local-first AI app" meant an Electron shell bundling a quantized model. What actually shipped is this: markdown scaffolding over a frontier-model CLI, where the "app" is prompts, file conventions, and a handful of small deterministic tools, and distribution is git clone. For developers building personal or niche software, that's a genuinely lower-friction model than standing up a SaaS — no hosting, no auth, no data-handling liability, and users who can read every instruction the agent follows before running it.

The costs of that model are real too, and this repo is a clean case study in all of them. You're locked to one vendor's harness — there's an AGENTS.md gesture toward portability, but the workflow leans on Claude Code's skills, subagents, and connectors. The toolchain ask is nontrivial: Python 3.10+, Bun, and a full LaTeX distribution with lualatex is a lot of yak-shaving for a non-developer, which tells you who this is really for. And the operating cost is invisible until it isn't — "no API keys" means the multi-agent review loops and PDF iteration cycles draw down your Claude Code plan, and the README doesn't attempt to quantify it.

The part that should make you pause

An agent that reads arbitrary web content and then writes documents sent under your name is a prompt-injection target, full stop. The repo handles this more honestly than most: postings are treated as untrusted input, the workflow follows no instructions embedded in them and fetches no links from their bodies, and the README concedes the defenses are "instruction-level, not a sandbox," advising you to skim outputs from unfamiliar boards before sending. That's the right disclosure. It's also an admission that the safety of the whole system rests on prompt discipline plus your attention — fine for a tool you supervise, worth remembering before you wire /gmail-sync and /notion-sync into the loop and start trusting it unattended.

There's also an arms-race question the repo can't answer. Recruiters are already drowning in AI-assisted applications, and screening tools on the hiring side are increasingly AI too. A tool that only tailors — grounded in your real profile, forbidden from fabricating credentials, with a human on the send button — is about the most defensible position in that race. But it accelerates the race all the same, and the equilibrium where every application is agent-polished mostly erodes the signal cover letters ever carried. Lorentzen told employers he was using it; that norm won't survive 12,000 forks.

Verdict

As a job-search tool, this is the best-engineered entry in a crowded, mostly scammy category — worth forking if you're a developer actively hunting, with the caveat that the results testimonial is a sample size of one. As a signal, it matters more: it's the clearest demonstration yet that an agent CLI plus a folder of markdown is a viable application platform, and that "fork it and own it" beats "sign up for our beta" for software this personal. Study the architecture even if you never run /apply.