ai-job-search: Claude Code Job Application Agent with Drafter-Reviewer

ai-job-search is an open-source Claude Code framework for AI-powered job applications. Built by Mads Lorentzen, a PhD geophysicist who lost his job and built this over three months, it uses a drafter-reviewer agent architecture where one Claude agent drafts a tailored CV and cover letter and a second Claude agent reviews, critiques, and refines the output before presenting it to the user. The framework ships with four slash commands: /setup builds a structured professional profile from CVs, LinkedIn exports, and reference documents; /scrape searches job portals, deduplicates, and ranks postings across five dimensions (skills, experience, cultural, location, career trajectory); /apply runs the drafter-reviewer pipeline to generate application materials; and /interview prep generates role-specific questions and talking points. The architecture treats job applications as adversarial documents — the reviewer agent looks for weaknesses a hiring manager or ATS would find, making the output significantly more effective than a single-pass AI generation.

BUSINESS PROBLEM

According to LinkedIn's Global Talent Trends (2025), the average job opening receives 250 applications, and recruiters spend an average of 7.4 seconds reviewing each resume. A developer spends 4-6 hours crafting a single tailored job application: researching the company, customizing their CV, writing a cover letter, and optimizing for ATS keyword matching. At a $75/hour loaded cost, that is $300-450 per application. With most developers needing 10-30 applications to land a role, the total time investment reaches 40-180 hours — an entire week of full-time work. Existing AI CV tools produce generic outputs that ATS systems flag as templated, and the most common failure mode is AI hallucinations that invent skills or achievements, which destroys credibility when discovered. The drafter-reviewer architecture solves this by having a second agent specifically audit for inaccuracies and ATS weaknesses.

WHO BENEFITS

For a developer in active job search. Situation: Spends 6 hours per application tailoring CV and writing cover letters. After 20 applications, that is 120 hours of unpaid work. Payoff: ai-job-search reduces application time to 45 minutes. The drafter-reviewer pipeline produces ATS-optimized, honest materials. Apply to more roles with better quality. For a career changer moving between industries. Situation: Current CV emphasizes experience in the old industry. Needs complete repositioning for a new field. Payoff: /setup builds a profile that highlights transferable skills. /apply generates materials that reframe experience for the target industry. The reviewer catches old-industry language that hiring managers would skip. For a PhD or academic transitioning to industry. Situation: Academic CV is 6+ pages with publications and teaching experience. Industry resumes should be 1-2 pages focused on impact. Payoff: ai-job-search extracts relevant industry experience from the academic record. The reviewer ensures output reads as industry professional, not academic.

HOW IT WORKS

Step 1. Fork the repository (1 min). Fork MadsLorentzen/ai-job-search on GitHub to your account. Clone locally. Step 2. Run /setup (30-45 min). Claude interviews you about background, skills, and career goals. Attach your CV, LinkedIn export, diplomas, and reference letters. Run /expand to crawl GitHub repos and portfolio sites for additional competencies. Step 3. Search jobs with /scrape (5 min). Claude searches job portals, deduplicates results, and ranks each posting against your profile across 5 dimensions. Use /add-portal for non-default job boards. Step 4. Apply with /apply (20 min). Claude drafts a tailored CV and cover letter. A second Claude agent reviews the output for weaknesses, inaccuracies, and ATS optimization gaps. Iterate until both agents agree. Step 5. Prep with /interview (10 min). Claude generates role-specific interview questions based on the job description and your profile. Includes talking points for likely technical and behavioral questions. Step 6. Track applications (optional). The framework maintains an application log with status, dates, and notes for each position. Export as CSV or Markdown.

TOOL INTEGRATION

TOOL: ai-job-search v1.0 (MIT, 19,500 GitHub stars). Role: Claude Code framework for AI-powered job applications with drafter-reviewer agent pattern. API access: github.com/MadsLorentzen/ai-job-search. Auth: Claude Code API key. Cost: Free, open-source. Gotcha: The /setup step requires 30-45 minutes of active engagement. Users who skip or rush the setup get low-quality outputs. The quality of applications directly correlates with setup thoroughness. TOOL: Claude Code (Anthropic). Role: AI coding agent that runs the ai-job-search framework. API access: claude.ai/download or npm. Auth: Claude subscription (Pro $20/mo or Max $200/mo). Cost: $20-200/month. Gotcha: The framework uses Claude Code's agent capabilities extensively. The Claude Pro plan provides sufficient usage for individual job seekers. Max plan recommended for heavy users running 10+ applications per week. TOOL: LinkedIn / Jobindex / Jobnet. Role: Job portal sources for /scrape command. API access: Public endpoints and unauthenticated page scraping. Auth: None for public pages. Cost: Free. Gotcha: LinkedIn's unauthenticated public endpoints have rate limits. If you hit rate limits, spread /scrape commands across multiple sessions or use /add-portal with additional sources.

ROI METRICS

Metric Before (Manual) After (ai-job-search) Source Time per application 4-6 hours 45 minutes Community estimate ATS pass-through rate ~30% (generic CV) ~60% (tailored) Community estimate CV iterations per role 3-5 manual edits 2-3 auto + 1 review Community estimate Interview prep time 2-3 hours 10 minutes Community estimate

The week-1 win: fork the repo, run /setup with your full background, and use /apply on one real job posting. Compare the drafter-reviewer output to your current CV. The strategic implication: adversarial agent architectures where one agent drafts and another critiques produce measurably better outputs than single-pass generation. This pattern applies beyond job applications to any AI-generated content that must survive human scrutiny.

CAVEATS

1. (moderate risk) Setup time investment: /setup requires 30-45 minutes of focused work. Rushing produces low-quality outputs. Mitigation: Treat setup as an investment. The quality of all future applications depends on this step.
2. (significant risk) AI hallucination: The framework may generate plausible-sounding but fabricated achievements if the profile is thin. The reviewer catches many but not all. Mitigation: Manually verify every generated application before submission. The /expand command helps surface real accomplishments.
3. (minor risk) Job portal coverage: Default portals are Denmark-focused (Jobindex, Jobnet, Akademikernes Jobbank). /add-portal generates search skills for any local job board. Mitigation: Use /add-portal for your region. LinkedIn search provides broad coverage across all countries.
4. (moderate risk) ATS detection: Some ATS systems detect AI-generated application materials. The drafter-reviewer architecture minimizes templated language but does not eliminate detection risk. Mitigation: Personalize the output with specific, non-AI-generated details about your interest in the company. The reviewer can help identify templated phrases.