claude-job-hunter

An autonomous job search agent that can prove what it sent.

It scouts 17 job board APIs, scores each posting against your CV, tailors that CV per posting against a fixed set of facts, submits through the ATS, and then reads the page back to confirm the application actually landed.

Runs on your Claude Code session. The skill workflow needs no extra keys; the standalone workers (scout, tailor, apply) want a GROQ_API_KEY or ANTHROPIC_API_KEY in workers/.env — see .env.example.

1,212 leads processed
  316 applications submitted autonomously
  248 verified on the target page after submitting

---

Why the verification matters

The aggregate confirmation rate was a healthy-looking 78%. Grading delivery per channel instead of in total told a different story:

Four integrations working perfectly and one failing silently for months. The average had been hiding it. That integration is now disabled.

Full write-up in docs/EVAL.md.

---

The CV is tailored, and it cannot lie

Every claim the model makes is checked against cv-facts.json before anything is rendered. Any id it invents is dropped, and any technology it names that is not in the profile replaces the summary with the base text.

This is not decoration. A prompt injection hidden inside a job description made the model write "10 years of Kubernetes" into the summary. The instructions in the system prompt did not stop it. The deterministic validator did.

That routing pattern was extracted into a standalone library: intent-gate.

---

Evals

A promptfoo suite runs over the CV tailoring stage. It grades the model and the validator separately, because a model that tries to fabricate and gets caught is a different outcome from one that succeeds:

cd workers/evals && npx promptfoo eval

The suite paid for itself when Groq deprecated a model: swapping to gpt-oss-120b moved summaryLength from failing 7/7 to passing, which revealed that the per-posting summary had been silently falling back to a generic one on every single application.

Details in docs/EVAL-CV-TAILOR.md.

---

Observability

Every LLM call is traced with the OpenTelemetry SDK using the GenAI semantic conventions, exported to a spans table in the same SQLite file.

model                  calls   avg latency   tokens   cost
openai/gpt-oss-120b        8        1187ms     9128   $0.0026

One tailored CV costs $0.0003 and takes 1.2 seconds. Local export rather than a hosted backend: answering "what did this run cost" should not require shipping your prompts to a third party.

---

Setup

git clone https://github.com/ale-aguirre/claude-job-hunter
cd claude-job-hunter/workers && npm install
# macOS / Linux
ln -s $(pwd)/.. ~/.claude/skills/job-hunter
# Windows (PowerShell, elevated)
# New-Item -ItemType SymbolicLink -Path "$env:USERPROFILE\.claude\skills\job-hunter" -Target (Resolve-Path ..)

Then, in Claude Code:

/job-hunter setup

The wizard asks six questions and writes profile.json.

The CV tailor and the eval suite read workers/cv-facts.json (your facts — gitignored). Without it they fall back to workers/cv-facts.example.json, a fictional candidate, so a fresh clone can run the tailor and the evals out of the box:

cd workers && npx promptfoo eval -c evals/promptfooconfig.yaml

To use your own facts: cp cv-facts.example.json cv-facts.json and edit it.

---

Commands

Standalone tools:

node workers/eval-report.mjs      # delivery rate per channel, orchestrator reliability
node workers/check-alive.mjs      # opens every pending posting, marks the dead ones
node dashboard/server2.mjs        # localhost:3000

---

Sources

LATAM: GetOnBrd, Torre, Workana, Computrabajo, Bumeran Remote in USD: Remotive, RemoteOK, We Work Remotely, Himalayas, Jobicy, Arc.dev ATS boards: Greenhouse, Lever, Ashby Other: HN "Who is hiring" monthly thread, X and Reddit hiring posts

Which ones run depends on your profession and work mode, from profile.json.

Postings go stale. check-alive.mjs opens each pending one and records whether it still exists: of 219 pending leads, 170 were live and 29 were gone.

---

Design notes

The model never decides what runs. It classifies; a registry maps that to an executor. A model returning a handler name turns a typo into a production failure.

Silence is not consent. Anything gated on human approval fails closed if the human does not answer.

A worker that fails three times gets paused, not retried forever, and resumes on the next clean cycle. One broken integration should not take down the run.

Costs are measured, not assumed. A cheap model handles classification, with automatic fallback on rate limits.

---

Stack

Node.js · Playwright · Claude API · Groq · SQLite · OpenTelemetry · promptfoo

MIT.