Jobsearch

/jobsearch — US + Danish Job Search + Application Assistant

Thin launcher over the job-portal CLIs in the ai-job-search repo. Default to the US sources unless the user asks for Denmark. Each CLI emits clean JSON to stdout; errors are JSON on stderr with exit 1. Same run rule everywhere: cd into the CLI's cli/ dir first (bun resolves node_modules by file location), and never bump @bunli/core past 0.7.0.

🇺🇸 US jobs — us-jobs CLI

cd ~/ai-job-search/.agents/skills/us-jobs/cli && bun run src/cli.ts <command> [flags]

Five sources, one normalized output shape: {meta, results:[{source,id,title,company,location,remote,salary,url,posted}]}. All accept --limit and --format json|table|plain.

Keyless — company job boards (work immediately, ✅ verified live)

For "jobs at ", the default is auto — it fans out across all three ATSes in parallel so you don't have to know which one the company uses:

bun run src/cli.ts auto <company> -q engineer --remote --limit 20   # tries greenhouse+lever+ashby, merges, dedupes

meta.sources_hit tells you which ATS actually had the company. -q filters by title text. Use the single-source commands only when you already know the ATS:

bun run src/cli.ts greenhouse <board>   --limit 20   # e.g. stripe, airbnb, coinbase
bun run src/cli.ts lever      <account>  --limit 20   # e.g. palantir
bun run src/cli.ts ashby      <board>    --remote     # e.g. Notion, Ramp, linear, openai

<company>/<board>/<account> is the company's slug in its careers URL (jobs.lever.co/<account>, jobs.ashbyhq.com/<board>, boards.greenhouse.io/<board>). NOT_FOUND from auto means the company isn't on any of the three keyless ATSes (try Adzuna, or check their careers page for a different ATS).

Federal — USAJOBS (✅ key configured, verified live)

bun run src/cli.ts usajobs -q "data scientist" --location "Austin, TX" --remote --limit 20

Broad private market — Adzuna (✅ key configured, verified live)

bun run src/cli.ts adzuna -q "software engineer" --where "Seattle, WA" --salary-min 120000 --remote --limit 20

Keys are already set up (2026-06-04) in us-jobs/cli/.env — bun auto-loads it, so commands just work from the cli dir with no export needed. The same keys are also in ~/.zshrc for interactive terminals. .env is gitignored (keys never commit). If you ever see a MISSING_KEY error, the .env was moved/lost — re-add ADZUNA_APP_ID/ADZUNA_APP_KEY (developer.adzuna.com) and USAJOBS_API_KEY/USAJOBS_EMAIL (developer.usajobs.gov). Never fabricate listings on a key error.

How to run a US search

1. Specific company → greenhouse/lever/ashby (no key, instant).
2. Federal/government → usajobs.
3. Broad "X jobs in Y city" → adzuna (needs key). Without an Adzuna key, fall back to querying a few relevant company boards, and tell the user a key unlocks market-wide search.
4. Run sources in parallel (separate Bash calls), dedupe by title+company, present a ranked shortlist (title — company — location — [remote] — salary — url).
5. Then offer /apply <url> to tailor a CV + cover letter for any posting.

🇩🇰 Danish jobs

Repo location

~/ai-job-search

If it's missing, clone it: git clone https://github.com/MadsLorentzen/ai-job-search ~/ai-job-search, then for each of the four CLIs run bun install (deps are pinned to @bunli/core@0.7.0 — do NOT bump to "latest", 0.8+ pulls a React/OpenTUI runtime that crashes headless).

CRITICAL run rule — cd into the CLI dir first

Bun resolves node_modules relative to the file's location. Always cd into the specific cli/ dir before running, or bun grabs a wrong global-cache version of @bunli/core and crashes with Cannot find module 'react/jsx-dev-runtime'.

cd ~/ai-job-search/.agents/skills/<cli>/cli && bun run src/cli.ts <command> [flags]

<cli> ∈ jobindex-search, jobnet-search, jobbank-search, jobdanmark-search.

All commands accept --format json|table|plain (default json). Pipe JSON to python3/jq to filter.

The four portals

The gated ones run, parse, and error-handle cleanly — they return data from the user's Mac (residential IP / logged-in), just not from headless cloud hosts. If a gated CLI returns empty/403/204, say so plainly; don't claim no jobs exist.

Command cheat-sheet

jobdanmark-search (✅ verified live)

cd ~/ai-job-search/.agents/skills/jobdanmark-search/cli
bun run src/cli.ts search --text "elektriker" --job-type fuldtid --municipality Odense --limit 10
bun run src/cli.ts search --category 227978 --job-type "fuldtid,deltid" --page 2
bun run src/cli.ts detail <slug>                 # slug from search results
bun run src/cli.ts categories                     # live job counts per category
bun run src/cli.ts autocomplete -q "it"           # job-title / category IDs
bun run src/cli.ts locations -q "Odense"          # municipality / zip / region

jobnet-search

cd ~/ai-job-search/.agents/skills/jobnet-search/cli
bun run src/cli.ts suggestions -q "syge"                                  # ✅ typeahead titles
bun run src/cli.ts occupations --search-string "sygeplejerske" --per-page 5  # ✅ ESCO codes
bun run src/cli.ts search --search-string "udvikler" --region HovedstadenOgBornholm --work-hours FullTime --per-page 10
bun run src/cli.ts detail <jobAdId>              # UUID from search

Regions: HovedstadenOgBornholm, Midtjylland, Syddanmark, OevrigeSjaelland, Nordjylland. Order: PublicationDate | BestMatch | ApplicationDate.

jobindex-search

cd ~/ai-job-search/.agents/skills/jobindex-search/cli
bun run src/cli.ts search -q "python" --jobage 7 --sort date --limit 10   # jobage: 1|7|14|30|9999
bun run src/cli.ts detail <id|url>

Area filtering isn't reliable via params — put the city in the query: -q "python aarhus".

jobbank-search (academic)

cd ~/ai-job-search/.agents/skills/jobbank-search/cli
bun run src/cli.ts search --key "data scientist" --location 2 --type 3 --limit 10
bun run src/cli.ts detail <id>

Multi-value filters are comma-separated: --type 3,6 --industry 10331. Code tables live in jobbank-search/cli/README.md (type/location/work-area/education/industry).

How to run a search

1. Resolve the user's intent → pick portal(s). Broad/any-sector → jobdanmark (most reliable). Academic → jobbank. Public-sector → jobnet. Largest private pool → jobindex.
2. For jobdanmark/jobnet, use autocomplete/occupations/locations first to turn free text into IDs when filtering precisely.
3. Run search with --limit to keep output small; parse the JSON and present a ranked, deduped shortlist (title — company — location — url).
4. Search several portals in parallel (separate Bash calls) when the user wants broad coverage, then dedupe by title+company across results.
5. If a portal returns empty/gated, name which one and why; offer to retry from the logged-in browser.

Full application workflow (the repo's own commands)

For end-to-end (fit evaluation → tailored LaTeX CV → cover letter → reviewer pass), the repo ships Claude Code slash commands in ~/ai-job-search/.claude/commands/:

• /setup — build the candidate profile (fills CLAUDE.md, profile files) from documents/, a pasted CV, or an interview.
• /scrape — multi-portal search ranked by fit (uses the CLIs above).
• /apply <url|text> — evaluate fit, draft CV (cv/main_<co>.tex, compile lualatex) + cover letter (cover_letters/, compile xelatex), reviewer critique, revise.

To use those, work inside ~/ai-job-search so the repo's CLAUDE.md profile + command files are in scope. The profile is currently a blank template — run /setup there first. Optional salary benchmarking: salary_lookup.py + tools/convert_salary_excel.py (needs a user-supplied salary_data.json).

Quality rules

• Never invent job listings, companies, salaries, or deadlines — every fact comes from CLI JSON output. If gated/empty, report that, don't fabricate.
• Keep Danish characters intact (the CLIs decode HTML entities; don't re-mangle æ/ø/å).
• Don't bump @bunli/core past 0.7.0 or remove the per-CLI tsconfig.json (typecheck OOMs without it).