job-hunt-agent

name	job-hunt-agent
description	Autonomous job-hunting pipeline for a software/AI fresher: audit + fix GitHub repos (CI, descriptions, flagship pinning), improve a LaTeX resume, search real remote jobs (India/global) via Jina+DDG when Exa is down, and run a SUPERVISED apply flow. Encodes the hard limits (no banned auto-apply bots, login wall is the user's, the browser_navigate session is invisible to the user) so the agent sets expectations correctly instead of looping on a dead end.
version	1.0.0
author	Hermes Agent
license	MIT
platforms	["windows","linux","macos"]

Job-Hunt Agent (fresher SWE / AI)

End-to-end workflow to make a fresher job-ready and surface real remote roles. Covers: GitHub audit/fix → resume (LaTeX) → remote job search → supervised apply.

Triggers

• "help me get a [high-salary / remote] job", "search jobs for me", "prepare my GitHub/resume for hiring", "apply for jobs automatically".
• User shares a GitHub profile and wants it job-ready.
• User wants an agent to "take over the browser and apply".

PHASE 0 — Set expectations FIRST (do this before any tool call)

The user will often demand a "fully autonomous agent that applies to everything". State the truth up front to avoid burning the session:

1. No banned auto-apply bot. Mass automated applying violates LinkedIn/Naukri/ Indeed/Wellfound ToS → account ban; ATS filters blast applications; it wastes a strong profile. Offer a supervised flow instead.
2. The browser I drive (browser_navigate) is a remote/headless session the user CANNOT see or type into. Never say "log in in my browser" as if they can — they can't. If login is needed, tell them to open their OWN browser.
3. The login wall (password / 2FA / CAPTCHA) is the user's step. The agent never types passwords or solves CAPTCHA (hard rule in computer-use skill too).

PHASE 1 — GitHub audit + fix (see github-job-readiness skill)

• Rank repos by stars/commits; pick 1–2 real, user-backed flagships and make them mandatory. For this user: Automated-Video-Generator (20★, TypeScript/Remotion) + sproutern-open-source (live Next.js, 367 tests green).
• Verify CI is green via API (actions/runs) AND check-runs (a green workflow can still show a red commit ✗ from one failing job — e.g. a flaky Render E2E). Fix by making heavy/flaky jobs non-blocking or skip-mode.
• Add a repo description (needs a token / user does it in UI — API 401 unauth).
• Rewrite profile README to a single lane + "🟢 Open to work".

PHASE 2 — Resume (LaTeX) improvement

See references/resume-improvement.md + templates/RESUME_improved.tex. Key fixes: "graduate" not "student"; add grad year + CGPA (if ≥7.5); lead with the two flagships and their proof (stars/commits/test counts); add OSCG mentor; align skills to the REAL stack; drop weak projects to hold one page. Keep the user's LaTeX style.

PHASE 3 — Remote job search (no login needed)

Use the Jina + DuckDuckGo HTML technique (Exa/mcporter is often broken here):

Q="remote AI ML engineer india fresher salary"
enc=$(python -c "import urllib.parse,sys;print(urllib.parse.quote(sys.argv[1]))" "$Q")
curl -s "https://r.jina.ai/https://html.duckduckgo.com/html/?q=$enc" \
  | grep -oE '\\[[^]]+\\]\\(https://duckduckgo.com/l/\\?uddg=[^)]+\\)'

⚠️ Only recommend boards that are FREE for candidates.

If a board charges candidates (even partially, e.g. salary hidden behind Premium, or "unlock applications" fees), DO NOT list it. Mark it clearly as PAID or skip entirely.

FREE boards (safe to recommend):

• Naukri — India's biggest board, free for candidates. Browse/apply without login.
• LinkedIn — Easy Apply is free. Needs login (user's step), but highest volume.
• Wellfound — free profile → apply to many startups, founder-direct hiring.
• RemoteAI — remoteai.io/v2/jobs/ai-ml-engineering (and NLP/Prompt Engineering)
• YC Work at a Startup — free, skill-based hiring at YC startups.

AVOID (PAID for candidates):

• RemoteOK — salary hidden behind Premium. Applying from there routes through paid layers. Do not recommend as an apply channel.

Save a curated JOB_SHORTLIST.md with direct links + a tailored cover note built from the user's flagships. Include only FREE boards.

PHASE 4 — Supervised apply (the only legit "auto-apply")

• Preferred: user applies manually from JOB_SHORTLIST.md (30 sec/role). Agent did the search + targeting + pitch.
• If user insists on agent-driven browser: enable computer-use (computer_use tool), launch the user's real Chrome, pause for the user to log in, then fill + submit one role at a time with the user watching.
• Pitfall — computer-use may not enumerate Chrome on Windows: if capture app="Google Chrome" returns 0 elements even though the window exists (cua-driver only sees the desktop layer), STOP debugging and fall back to manual apply. The login wall is the user's step anyway, so the automation gain is small.

Pitfalls

• Only recommend FREE job boards. Never suggest boards where candidates need to pay/premium for salary info or applying (RemoteOK, Premium-locked boards, "resume unlock" fee schemes). Free only: Naukri, LinkedIn, Wellfound, RemoteAI, YC Work at a Startup.
• Don't loop on making browser_navigate "visible" to the user — it's headless.
• Don't promise "high salary" to a fresher beyond market reality (India AI/ML fresher ₹6–12 LPA; anything >₹20 LPA "no experience" is a scam).
• Don't paste a password into chat to "log in for them" — security + ban risk.
• Exa/mcporter is unreliable here; default to Jina+DDG, not a failed retry loop.

References

• references/remote-job-search.md — Jina+DDG recipe, board list, salary reality.
• references/resume-improvement.md — fresher LaTeX resume fix checklist.
• templates/RESUME_improved.tex — copyable improved resume template.