ai-job-search

ai-job-search (Mads Lorentzen, MIT) is “an AI-powered job application framework built on Claude Code. Fork it, fill in your profile, and let Claude evaluate jobs, tailor CVs, write cover letters, and prepare you for interviews.” A fork-and-customize repo that turns Claude Code into a full job-application assistant — and a clean example of this wiki’s skills + harness + orchestration thesis applied to a non-coding vertical.

Built on Claude Code’s native primitives

Slash commands (.claude/commands/) as the UI: /setup (ingest CV PDFs / LinkedIn / diplomas / guided interview), /scrape (rank jobs across Danish portals), /apply <url> (the orchestrator), /expand, /upskill, /reset.
Skills (.claude/skills/) as modular logic: a core job-application-assistant (profile, behavioral assessment, writing-style, job-evaluation framework, LaTeX templates, STAR examples), plus job-scraper and upskill. .agents/skills/ holds TypeScript/Bun Danish-portal CLI tools.
CLAUDE.md as the master profile + workflow rules.
No MCP/webhooks — relies on Claude Code’s native agent dispatch + filesystem.

The `/apply` drafter–reviewer pipeline

Parse posting → score fit → draft CV + cover letter in LaTeX → spawn a reviewer subagent that independently researches the company and critiques the draft → drafter revises → compile + visually inspect PDFs → iterate to exact length → present with a checklist. The drafter/reviewer split is a concrete instance of agent-orchestration‘s adversarial verification and the two-agent review of agents-never-do-alone (“eliminating single-pass blind spots”).

Notable design patterns

Output-grounded verification loop — it does not trust LaTeX to render correctly: it compiles (lualatex/xelatex), reads the rendered pages, and applies targeted fixes (\needspace, \enlargethispage) until visual inspection passes (CV exactly 2 pp, letter 1 p). Verifying the artifact, not the generation — a reliability discipline like the harness/durability threads. watch-skill generalizes this move to video: read the rendered PDF here, watch the rendered screen-recording there — the same verify-the-artifact loop in a modality text checks can’t reach.
Relevance-weighted CV cutting — on overflow it scores each line by job-relevance, uniqueness, and cover-letter dependency and cuts the lowest-value line, not the oldest.
Token-efficient reviewer dispatch — reviewer gets drafts inline; checklist runs once.
Latent-opportunity discovery — ranks by work actually performed, not job titles.

Significance

Most sources here are coding agents or general frameworks; ai-job-search shows the same skills+harness+subagent pattern generalizing to an everyday non-coding vertical (job hunting), language/country-agnostic. Credits Anthropic’s Claude Code and Mikkel Krogholm’s Danish job CLI skills. Stack: TypeScript 40% / Python 34% / TeX 25%; needs Claude Code CLI, Python 3.10+, Bun, TeX Live (lualatex + xelatex).

Re-verified 2026-08-08 (staleness queue)

The page was flagged volatile-past-window at 61 days and the re-read changed its standing. It is no longer a personal repo with a niche scraper: 30.8k stars and 10.5k forks, MIT, actively maintained, with a “Community forks & adaptations” thread documenting regional re-implementations — so the Denmark-focused caveat below now describes the upstream rather than the thing people run.

The command surface has grown past application drafting into pipeline management: /interview, /outcome, /rank, /upskill (skill-gap analysis with learning plans), /html-report, /notion-sync and /gmail-sync (status detection from mail) join the original /setup, /scrape, /apply. That is the accumulating-state shape arriving in a vertical that started as one-shot document generation, and the Gmail/Notion syncs make it the corpus’s clearest case of an agent skill-pack growing a system of record around itself.

The author now states explicitly that the project “is not affiliated with, endorsed by, sponsored by, or maintained by Anthropic” — worth recording because the fork-a-repo distribution model puts a vendor’s name on work the vendor does not own.

Caveat

A personal MIT repo / README — design described by its author, no independent evaluation; the upstream portal scraping is Denmark-focused, and the fork count is adoption, not evidence it works.

agentic-coding-harness · agent-skills · agentskills-spec · agent-orchestration · agents-never-do-alone · claude-code-best-practices · anthropic