JobSearch Agent
TMIB/JobSearch
Pipeline automático de busca de emprego com agentes de IA especializados
Ver no GitHubSistema de busca de emprego automatizado que executa diariamente, encontra vagas relevantes, as avalia com raciocínio de IA, gera currículos personalizados e notifica quando há novos leads prontos. Funciona com quatro agentes especializados (Search, Evaluation, Resume Tailoring e Orchestrator) que se integram em um pipeline tunable interativamente via Claude Code.
Instalação
npx convez-ai TMIB/JobSearch1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar busca diária de vagas em múltiplos sites de recrutamento
- Avaliar e filtrar vagas com raciocínio de IA (não apenas palavras-chave)
- Gerar currículos personalizados para cada vaga descoberta
- Calibrar agentes de IA com feedback iterativo sobre avaliações incorretas
Veredito

Projeto robusto e bem arquitetado para profissionais que dominam shell scripting e integração de APIs. Requer conhecimento técnico significativo (Claude Code CLI, SerpAPI, Python, Bash) e disposição para tuning iterativo dos agentes — não é plug-and-play. Vale a pena para quem busca refinar sistematicamente sua pipeline de recrutamento.
README
A tradução para o português deste README ainda está na fila.
Agentic Job Search Pipeline
An AI-powered job search system that runs daily, finds relevant listings, evaluates them with reasoning (not keyword matching), generates tailored resumes, and notifies you when new leads are ready.
What This Is
This is a starting point for building your own automated job search pipeline using Claude Code. It's not a turnkey app you install and run — it's a set of agent prompts, config files, and scripts that you customize to your career profile, then use Claude Code interactively to tune and improve over time.
The key idea: Four specialized AI agents work together in a pipeline. Each has its own prompt file and feedback file. You review the output, tell Claude what it got wrong, and the system gets smarter with each run. After a few days of tuning, it produces leads and resumes that genuinely match what you're looking for.
This tool is designed to be used with Claude Code in an interactive conversational manner. You'll work with Claude to:
- Fill in your career profile and experience inventory
- Review generated resumes and flag issues ("don't lead with that bullet for AI roles," "that company is actually on-site, not remote")
- Add calibration examples when the scoring is off
- Generate cover letters for specific roles
- Process individual listings you find yourself (
--urlmode)
The automation handles the daily grind of searching and filtering. The interactive Claude sessions handle the judgment calls.
How It Works
Four specialized agents, each with its own tunable prompt:
-
Search Agent (Sonnet) — Queries SerpAPI Google Jobs API, which aggregates listings from Indeed, Glassdoor, LinkedIn, Ladders, ZipRecruiter, and dozens more. Returns full job descriptions, salary data, and application links. Rotates through your search queries over a 4-5 day cycle.
-
Evaluation Agent (Sonnet) — Scores each listing on 6 weighted dimensions: level match, role category, location fit, coding interview risk, narrative fit, and company maturity. Uses AI reasoning to distinguish "Director of QA Engineering" from "QA Engineer," detect fake Director titles at tiny startups, and assess coding interview risk from company culture signals. Has a feedback file where your corrections accumulate, so it calibrates over time.
-
Resume Tailoring Agent (Opus) — Takes above-threshold leads and generates a customized HTML resume aligned to each listing's keywords and requirements. Uses an experience inventory that tracks what keywords are safe vs. what would overstate your experience. Builds a library of approved resumes and reuses them as starting points for similar future roles.
-
Orchestrator — Shell script that coordinates the pipeline, deduplicates against previous runs, converts resumes to PDF via headless Chrome, generates an HTML report with color-coded scores and clickable links, and sends a desktop notification.
Getting Started
Prerequisites
- Claude Code CLI installed and authenticated
- SerpAPI account (free tier: 250 searches/month, or $25/month for 1,000)
- Google Chrome (for headless PDF generation — auto-detected on macOS, Linux, and WSL)
- Python 3 (for report generation and JSON parsing)
- Bash (macOS/Linux natively, Windows via Git Bash or WSL)
Step 1: Clone the project
git clone https://github.com/TMIB/JobSearch.git
cd JobSearch
All script paths are relative to the project root — no path configuration needed. Clone it wherever you like.
Step 2: Add your SerpAPI key
Sign up at serpapi.com (free tier, no credit card). Copy your API key from the dashboard and add it to:
automation/run_search.sh— replaceYOUR_SERPAPI_KEY_HEREautomation/serpapi_search.sh— replaceYOUR_API_KEY_HERE
Step 3: Configure your career profile
This is the most important step. Open Claude Code in this project directory and work with Claude to fill in:
CLAUDE.md— Your career background, target roles, confirmed experience details, and critical rules (what to never fabricate, what titles are accurate, etc.)career/career_strategy_guide.md— Your situation assessment, target role categories, external narrative, and constraintsautomation/experience_inventory.yaml— Your skills with safe/avoid keywords, certifications, industry preferences, and per-role-category guidanceautomation/eval_agent_prompt.md— The "Who [CANDIDATE] Is" section needs your profile
Tell Claude about your background and let it help you structure these files. The more honest and detailed you are here, the better the output.
Step 4: Customize your search queries
Edit queries.txt with searches tailored to your target roles, industries, and locations. The search agent rotates through ~12 queries per day, so include enough to cover your target categories.
Step 5: Create your first resume template
Create at least one hand-crafted HTML resume in a leads/ subfolder. This becomes the template the resume agent uses as a starting point. Look at the HTML/CSS conventions in CLAUDE.md for the expected format (Georgia serif font, letter-size margins, break-inside:avoid, etc.).
Place your current resume PDF in the project root as a reference.
Step 6: Test run
# Search only — verify SerpAPI works and listings come back
./automation/run_search.sh --search-only
# Full pipeline — search, evaluate, generate resumes + PDFs
./automation/run_search.sh
# Process a specific listing URL you found yourself
./automation/run_search.sh --url "https://example.com/job/12345"
# Skip evaluation for a URL (you already know it's a good fit)
./automation/run_search.sh --url "https://example.com/job/12345" --skip-eval
# Recover a run that died after search (e.g. an eval timeout) — reuses the
# existing search results and restarts at evaluation, no re-search / no re-spend
./automation/run_search.sh --skip-search
# Force a full run even if today's report already exists
./automation/run_search.sh --force
After the first run, open leads/new_leads_report.html in your browser and review the results with Claude. Flag what's wrong — this is where the tuning begins.
Step 7: Schedule daily runs
macOS (launchd):
Edit automation/com.jobsearch.plist — replace /path/to/your/project with your actual project path (launchd requires absolute paths) and set your preferred hour. Then:
cp automation/com.jobsearch.plist ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.jobsearch.plist
Important: launchd loads a copy of the plist into memory. If you later change the schedule or paths in the project's plist file, you must copy and reload:
launchctl unload ~/Library/LaunchAgents/com.jobsearch.plist
cp automation/com.jobsearch.plist ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.jobsearch.plist
Editing the file in ~/Library/LaunchAgents/ directly also works, but you still need to unload/load for changes to take effect.
Linux (cron):
crontab -e
# Add this line (adjust the path):
0 7 * * * /path/to/your/project/automation/run_search.sh >> /path/to/your/project/automation/logs/cron.log 2>&1
Windows (Task Scheduler): Create a scheduled task that runs daily and executes:
bash C:\path\to\your\project\automation\run_search.sh
Requires Git Bash or WSL.
Runs daily at 7 AM. Edit the plist to change the time.
macOS sleep gotcha (important): launchd fires scheduled jobs while the Mac is in DarkWake (maintenance sleep), where it won't stay awake long enough for the multi-minute evaluation API call — the connection stalls and the run times out before producing a report. The shipped plist mitigates this two ways: it wraps the job in caffeinate (keeps the Mac awake for the whole run) and sets RunAtLoad (a run missed while the Mac was fully asleep/off is caught up at next login/boot; the script's catch-up guard skips it if today's report already exists). For maximum reliability on a laptop, also schedule a power-on wake a couple minutes before your run time (one-time, needs sudo):
sudo pmset repeat wakeorpoweron MTWRFSU 06:58:00 # ~2 min before a 7 AM run
If a scheduled run ever still fails partway, recover it without re-searching: ./automation/run_search.sh --skip-search.
Daily Workflow
- Pipeline runs automatically, you get a macOS notification
- Open
leads/new_leads_report.htmlin your browser - Review leads — each card has score badges, apply links, and local file links to the listing, resume, and PDF
- Open Claude Code and work through the results:
- "This role is actually on-site in Ohio, not remote" → Claude logs it to eval feedback
- "The resume for Acme Corp has a bad page break" → Claude fixes it and logs the pattern
- "I applied for the Acme Corp role" → Claude tracks it in applications.yaml
- "I need a cover letter for this one" → Claude generates it
- Each correction makes the next run more accurate
Tuning the System
The system improves through three feedback loops:
| Agent | Config | Feedback | What You Tune |
|---|---|---|---|
| Search | queries.txt | — | Which queries, platforms |
| Evaluation | search_config.yaml | automation/eval_feedback.yaml | Scoring weights, thresholds, calibration examples |
| Resume | automation/experience_inventory.yaml | automation/resume_feedback.yaml | Keywords, emphasis, quality corrections |
Common adjustments after the first few runs:
- Too many false positives? Raise
generate_resumethreshold insearch_config.yaml - On-site roles scoring too high? The eval prompt has a location score cap — adjust it
- Wrong role types getting through? Add dealbreakers to
search_config.yaml - Resume quality issues? Add corrections to
resume_feedback.yamlwith severity levels - Scoring a role type wrong? Add a calibration example to
eval_feedback.yaml
Lessons Learned
These tips come from real-world usage of this pipeline:
- SerpAPI location data is unreliable. "Remote (Anywhere)" sometimes means hybrid on-site. The eval agent should read the full description for on-site signals, not trust metadata.
- Aggregator sites (Virtual Vocations, BeBee, talent.com) often have stale listings. Flag them with higher stale risk.
- "Director" at a 10-person startup with 3 IC reports is not a real Director. Teach the eval agent to check team size, reporting structure, and company size.
- Hourly rates at staffing firms signal contract roles, not FTE. Convert to annual and compare to your floor.
- Location-based pay bands matter. Check your city's band, not the SF/NYC number.
- The resume agent will create too many full job sections and break page layout. Budget 4-5 full sections max, fold the rest into single-line entries in Additional Experience.
- Never use internal tool names or codenames in external documents.
- The first run is the noisiest. After 2-3 days of feedback, the quality improves dramatically.
Cost
- SerpAPI: Free tier (250/month) or $25/month for 1,000 searches
- Claude API: ~$3-8/day depending on how many resumes are generated
- Search (Sonnet): ~$0.30/run
- Evaluation (Sonnet): ~$0.20/run
- Resume (Opus): ~$1.50 per resume
- Budget cap: Configurable in
search_config.yaml(recommend $20/run during initial tuning, lower to $5-10 after)
File Structure
project/
├── README.md # This file
├── CLAUDE.md # Your career profile and rules (customize)
├── queries.txt # Search queries (customize)
├── search_config.yaml # Scoring config (mostly ready, tune over time)
├── career/
│ └── career_strategy_guide.md # Career strategy (customize)
├── automation/
│ ├── search_agent_prompt.md # Search agent instructions
│ ├── eval_agent_prompt.md # Eval agent instructions (customize profile section)
│ ├── eval_feedback.yaml # Your eval corrections (grows over time)
│ ├── resume_agent_prompt.md # Resume agent instructions (customize experience section)
│ ├── resume_feedback.yaml # Your resume corrections (grows over time)
│ ├── experience_inventory.yaml # Your skills/keywords reference (customize)
│ ├── serpapi_search.sh # SerpAPI helper (add API key)
│ ├── run_search.sh # Main orchestrator (update paths + API key)
│ ├── generate_html_report.py # HTML report generator
│ ├── com.jobsearch.plist # launchd schedule (update paths)
│ └── logs/ # Run logs (auto-created)
└── leads/
├── applications.yaml # Application tracker
├── seen_listings.jsonl # Dedup tracker (auto-created)
├── new_leads_report.html # Daily HTML report (auto-created)
└── {CompanyName}/ # One folder per lead (auto-created)
├── {CompanyName}.txt # Job listing text
├── fit_analysis.md # Evaluation reasoning + resume strategy
├── resume_*.html # Tailored resume
└── [Name] - Resume.pdf # Auto-generated PDF
Credits
Built with Claude Code, April 2026.