AI Job Search, an open-source project built by Danish geophysicist Mads Lorentzen, has reached 34,000 stars and close to 12,000 forks on GitHub in just five months, according to BlockTempo. Lorentzen created the framework after his position was cut at the end of 2025, turning his job-hunting process into a specification that Claude Code could run. He said the workflow helped him send 69 tailored applications, land 20 first-round interviews, and return to work as an AI engineer.
He later released the full package as open source under the name AI Job Search.
A software project with very little traditional code
The repository contains 276 files, but most of them are Markdown rather than application code.
Its core workflow is organized through 13 slash commands. Those commands cover profile setup through /setup, job scraping through /scrape, batch scoring and ranking through /rank, tailored resume and cover-letter generation through /apply, interview preparation through /interview, and result logging and archiving through /outcome.
Another nine Markdown files hold the knowledge layer behind the process, including candidate background, behavioral traits, writing style, a fit-scoring framework, and STAR interview examples.
The parts that count as actual software are limited to two groups: six job-site scraping tools written in TypeScript and two LaTeX resume templates. Users can fork the project to a local machine, fill in their own information, and then let Anthropic's Claude Code run the workflow.
The /apply flow is split into eight steps
BlockTempo highlighted /apply as the most notable piece of the system.
Instead of throwing a job posting at a model and asking for a resume in one shot, the framework breaks the work into eight stages: parsing the job description, scoring fit against the user's profile, drafting the resume and cover letter, launching a fresh-context review agent to research the company and criticize the draft, rewriting based on that critique, compiling the output into PDF and having the model inspect the rendered result, extracting the PDF text layer for an ATS check, and producing a validation checklist.
Two features stand out: PDF visual feedback and ATS validation
The report said two parts of the process are where the real differentiation shows up.
The first is PDF visual feedback. A common LaTeX resume problem is that the source file looks fine while the rendered PDF breaks in subtle ways. A job title may get pushed onto the next page as a widow line, or bullet-point fonts may silently fall back to the body default. This workflow forces a compilation step, then has the model read the rendered pages and revise the file until the resume is exactly two pages and the cover letter is exactly one page.
The second is ATS checking. ATS, or applicant tracking systems, read the embedded text layer in a PDF rather than the layout a human sees on screen. BlockTempo noted that LaTeX can produce files that look polished visually but extract as garbled text, such as an email address turning into icon characters or a two-column layout scrambling line order. The workflow checks the text layer directly and compares it with keyword coverage from the job posting.
The project also hard-codes one rule: if a skill is not present in the user's actual experience, it must be marked as a gap. The system does not allow keyword stuffing, and every claim has to match the user's real background.
Usage limits are built into the project
AI Job Search comes with six built-in job search tools, but four of them are aimed at Denmark-based sites. Users in other regions need to create their own through /add-portal.
Lorentzen also uses a warning box in the project documentation to tell users to keep the repository private if they are using it for themselves, and to fork only when contributing code back.
Hiring teams are also adopting AI
The report places the project in a wider hiring trend. Research from Stanford HAI found systemic inequality in AI hiring tools affecting Black and Asian candidates. Another study found that AI screeners can recognize their “own handwriting,” with resumes written by the same model seeing a 60% higher acceptance rate.
BlockTempo also linked to related coverage on bias in AI hiring tools and on the rise of late-night remote interviews.