convez.ai
AgenteAutomaçãoIntermediário
a-n-jelly

CareerScout

a-n-jelly/careerscout

Agente para gestão completa de busca de emprego com IA

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent Claude que automatiza a busca de emprego end-to-end: encontra vagas diárias pontuadas, avalia compatibilidade com sua carreira, personaliza resume e carta de apresentação, e rastreia candidaturas. Baseado em loops de feed (descoberta) e aplicação (execução).

Instalação

$npx convez-ai a-n-jelly/careerscout

1 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

Estrelas
0
Forks
1
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automatizar descoberta de vagas alinhadas ao seu perfil
  • Gerar resumes e cartas personalizadas por vaga
  • Rastrear pipeline completo de candidaturas e feedbacks
  • Calibrar scoring de vagas com base em rejeições e aceitações

Veredito

Agent maduro e bem estruturado para quem busca sistematizar a procura de emprego. Exige setup Python e manutenção de queries de scraping, mas o loop de calibração aprende com feedback do usuário. Ideal para candidatos em busca ativa que querem gastar menos energia em repetição manual.

README

A tradução para o português deste README ainda está na fila.

CareerScout

A Claude Code agent for your job search. Paste your resume once — it finds roles, scores fit, tailors your resume, writes cover letters, and tracks applications.

You don't need to memorise commands. Paste a job description and it runs /assess. Paste recruiter feedback and it files it. Mention an outcome ("got rejected", "interview tomorrow") and it logs it and tells you what to do next. Type /help any time for a command reference.


How it works

Two loops: the feed finds roles worth applying to, the application loop helps you apply well.

Feed loop

fetch.py runs daily (or manually) and scrapes LinkedIn, Indeed, and direct ATS boards for roles matching your criteria. It writes raw results to today_raw.json. filter_roles.py then removes dismissed roles and out-of-scope roles using your personal rules from feed-agent/my_filters.json (generated by /feed-setup), writing the survivors to today.json. Run /feed in Claude Code to score and rank them against your profile, producing a shortlist in feed/daily-feed-output.md. React to the results with /calibrate and the scoring gets sharper over time.

CommandWhat it does
/feedScore today's fetched roles and write the shortlist
/calibrateTune scoring from your feed reactions — updates sources.md, differentiators.md, and learnings.md
/feed rejectLog why a role wasn't right — teaches the feed to stop surfacing similar ones
/retroPeriodic review — rejection patterns, pipeline health

Scoring calibration: The first few feeds will have noise. Run /calibrate after each early run until the scoring feels right — usually 3-5 runs.

Fetch calibration: /calibrate only tunes how roles are scored — it doesn't affect which roles are fetched in the first place. If you see a role on LinkedIn that never appeared in the feed, that's a fetch gap: the scraper either missed it (JobSpy returns the top N results per query, not a complete list), a query term doesn't cover it, or an ATS slug is broken. Fetch calibration is an ongoing process — if you notice a pattern, add or refine queries in context/sources.md.

Application loop

Paste a job description and /assess gives you a fit verdict, ranked recruiter concerns, and a cover letter angle. From there, /tailor-resume adjusts your resume bullets for the role and /cover-letter writes the letter in your voice. Every outcome — interview, rejection, offer — gets logged to your pipeline so nothing falls through.

CommandWhat it does
/improve-resumeReview and strengthen your master resume
/assessFit verdict, recruiter concerns, and culture fit check — feedback updates your context files
/tailor-resumeTailor your resume to the JD
/cover-letterWrite a cover letter in your voice
/notesCapture recruiter feedback or call notes
/updateLog an event — interview, rejection, offer
/trackView and manage your full pipeline

Setup

Prerequisites: Claude Code installed. Python 3.x + python-jobspy only needed for the job feed — you can skip that and still use everything else.

1. Clone and open

git clone https://github.com/a-n-jelly/CareerScout.git
cd CareerScout

Open Claude Code in this directory.

2. Install Python dependencies (feed only)

Skip this if you're not using the job feed. Otherwise:

python3 -m venv feed-agent/.venv
feed-agent/.venv/bin/pip install -r feed-agent/requirements.txt

Or with system Python: pip install python-jobspy.

3. Run /setup

Type /setup in Claude Code. It walks you through everything in one flow:

  • Saves your resume
  • Creates your context files from templates
  • Calibrates your writing voice (optional)
  • Configures the job feed (optional)

Each step can be skipped and done later.

macOS: if you use the daily scheduler, macOS will eventually show a dialog saying "bash wants to access files in your Documents folder." That's the job feed — click Allow. If you accidentally denied it: System Settings → Privacy & Security → Full Disk Access → add /bin/bash (use ⌘⇧G to navigate there) and toggle it on.


Context files

The agent reads context/ at session start — you don't re-explain your background each time.

FileWhat it isAuto-updated?
resume.mdMaster resumeNo
profile.mdOff-resume context, Mnookin preferencesYes — by /feed-setup (Mnookin doc), /assess (Mnookin calibration)
bank.mdCover letter style, talking points, Q&A, assess calibrationYes — by /voice-setup, /assess, /tailor-resume
state.mdLive application pipelineYes
sources.mdFeed criteria: roles, locations, comp, deal-breakersYes — by /calibrate
differentiators.mdYour edges and hard gaps — used in feed scoringYes — by /assess and /calibrate

What's next

The feed is the area with the most room to grow. In priority order:

  • Workday ATS support — many companies use Workday for hiring. Direct ATS scraping for Workday would close the biggest remaining fetch gap.
  • Wellfound as a source — startup and early-stage roles that don't appear on LinkedIn. Worth adding once core calibration is stable.
  • Calibration improvements — smarter feedback loops: domain rejections updating Domain Preferences automatically, pattern summaries after enough runs, Edge scoring feedback when a strong-match role gets rejected.
  • Feed debug command — paste a LinkedIn URL to diagnose why a role wasn't fetched (wrong query, broken ATS slug, outside result window).

Contributing

The agent's behaviour is defined in .claude/commands/ (command protocols) and .claude/skills/ (reusable skill instructions). CLAUDE.md has the full file index and session rules. The feed pipeline lives in feed-agent/ — fetch.py, filter_roles.py, and enrich.py are the three steps that run before Claude scores anything.

PRs welcome. If you're adding a new command, follow the pattern in .claude/commands/assess.md — explicit protocol steps, no ambient assumptions.

Ver o original no GitHub

Bom para

Relacionados