ai-job-search: Claude Code Job Hunt Agent on Local

Background and Context

Job hunting is repetitive, fiddly and draining. For every posting you read the description, judge your fit, rewrite a résumé, draft a cover letter and, if you get that far, prepare for the interview. Most people either grind through it by hand or hand their documents to an online résumé optimizer. Those services are usually black boxes, and they require you to upload personal data to someone else's server. ai-job-search takes a different route. It is an open-source job-hunt framework built on Claude Code, and the whole workflow runs on your own machine. You fork the repository, fill in a personal profile, and end up with a job-search workbench that you are free to modify.

The author, Mads Lorentzen, is a geophysicist by training. After his position was cut in late 2025, he built the tool for himself. According to the README, he used the same /scrape, /apply and /interview flow to send sixty-nine tailored applications, earned twenty first-round interviews, signed one contract, and started work as an AI engineer in June 2026. The repository has passed fifty thousand stars on GitHub, which suggests real demand for a job assistant that the user controls.

What is ai-job-search, and how does it differ from a generic agent framework? It is not a general-purpose agent framework. It uses a command-line AI assistant, Claude Code, as the runtime for one vertical workflow. Slash commands and skill files carry the domain logic, so the project is a clear example of building a specialized assistant on top of an existing agent tool rather than writing an orchestration layer from scratch.

Deep Analysis

The capabilities fall into three layers. The first is a self-portrait. The /setup command walks you through a personal profile and turns your experience, skills and preferences into structured files. Every later step treats those files as the source of truth, which limits the risk that the model invents details about your career. The second layer is discovery and evaluation. The /scrape command searches job portals and returns a list of postings with match ratings. You pick the ones you want and pass them to /apply. That command first assesses the fit, assigns a score and says whether the role is worth pursuing, and only then moves into writing. The third layer is the drafter-reviewer pipeline, and it is the most distinctive design choice. One agent drafts a tailored résumé and cover letter in LaTeX. A separate reviewer agent critiques the draft, and the draft is then revised against that critique before the final version is produced. Splitting the writer from the critic makes empty phrasing and weak matches easier to catch than asking a single model to write everything in one pass. The framework also builds in career-coaching practice: structured evaluation criteria, a forward-looking way of framing the cover letter, and an optional salary benchmark.

The README is explicit that the core flow works in any language and any country. The job-search skill, however, was written for the Danish market and covers portals such as Jobindex, Jobnet and Akademikernes Jobbank. By design, it can be replaced with a local job board in your own country. The entry barrier is modest, but the prerequisites should be clear. You need the Claude Code command-line tool, which has no free tier. It requires a Claude Pro, Max or Team subscription, or Anthropic API credit; for occasional use, paying per token is often cheaper. You also need Python 3.10 or later and Bun, which runs the job-search command-line tool. A typical session looks like this: fork the repo, start Claude Code in the directory, run /setup, browse new postings with /scrape each week, run /apply on the roles you like to get a résumé and cover letter, and use /interview to prepare before a conversation. On documentation, the README draws a flow diagram that shows how the three main commands relate. The repository also provides an AGENTS.md file for people who use other agent tools such as Codex, Antigravity or Gemini CLI. The job-search skill works directly in those tools, and community forks share adaptations of the full flow in the discussions. A CI badge indicates that the author maintains quality. The project accepts Ko-fi donations and GitHub contributions, and it states that it has no affiliated cryptocurrency, token or paid sponsorship program, warning readers about impersonation scams.

Industry Impact

Several points stand out for the industry. First, the project shows that Claude Code can serve as a base for personal workflows, not only for writing software. Slash commands plus skill files are enough to assemble a complete domain assistant. Developers who want to build similar vertical tools now have a concrete sample to study.

Second, the local-first, fork-to-own approach lets users keep control of their data and their prompts. They do not depend on the pricing or policies of a hosted platform. For an activity as personal as job hunting, where a profile contains employment history and salary expectations, that control is a real selling point. Third, the author says he told interviewers that he used the tool. By his account, this often led to genuine technical conversations rather than suspicion. That offers a useful example of transparency in AI-assisted job applications. The risks are equally clear. Generated documents still need to be checked line by line by the applicant, because any exaggeration will surface in an interview. Sending many tailored applications creates ongoing model costs. The job-search skill depends on the page structure of specific portals, so a site redesign can break it. And because the framework is tied to Claude Code, moving to another tool takes extra adaptation work.

Outlook

Three things are worth watching. The first is whether the community adapts the job-search skill to more national job boards, since the Denmark-specific scraper is the main barrier for anyone outside that market. The second is whether ports to other agent tools mature, given that AGENTS.md already points in that direction and forks are sharing adaptations of the full flow.

The third is whether the evaluation and review stages can produce more stable quality standards over time. The scoring in /apply and the critique from the reviewer agent are the parts that decide whether a draft is actually worth sending. If contributors refine those criteria and share what works, the project could move from one person's successful workflow to a reusable reference for AI-assisted applications. For now, ai-job-search is best read as a well-documented case: one engineer, one agent tool, and a pipeline that stayed on his own computer from the first search to the signed offer.