Resume Builder - Customizador de Currículo com IA
Khushir474/Resume-Builder
Adapta currículo a vagas com pipeline multi-etapa e validação por council
Ver no GitHubSkill do Claude Code que adapta currículo para vagas específicas, otimizando para ATS e aplicando classificação de cargo automática. Funciona em modo interativo via comando ou automático ao detectar novos arquivos de job description, com pipeline de validação em camadas para reduzir alucinações.
Instalação
npx convez-ai Khushir474/Resume-Builder1 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
- Candidatos gerando múltiplos currículos otimizados para diferentes vagas
- Automação de busca de emprego com processamento de job descriptions em lote
- Validação de keywords e skills relevantes para cada posição
Veredito

Ferramenta sólida para quem busca automatizar customização de currículo com rigor técnico — councils de validação (classification + keywords) reduzem hallucinations reais. Requer setup em macOS com pdflatex e estrutura de pastas específica; ideal para candidatos em processo ativo de recrutamento em escala.
README
A tradução para o português deste README ainda está na fila.
Resume Builder
A Claude Code skill that tailors a resume to a job description — ATS-optimized, role-classified, and compiled to PDF. Runs interactively via /resume-builder or fully automatically when a new JD file is dropped into the vault.
Modes
| Mode | Trigger | Gates |
|---|---|---|
| Interactive | /resume-builder or /resume-builder path/to/jd.md | Role type + keyword review (manual) |
| Auto | Drop .md into ~/JobSearch/JDs/ | Council-gated — auto or notify based on confidence |
| Compile | /resume-builder --compile [RoleType] | None — reads past applications, writes archetype |
Auto mode pipeline
JD file dropped into ~/JobSearch/JDs/
│
▼
Parse frontmatter
(company, role, date)
┌─ Company extraction ────────────────────────────────┐
│ title: "Role at Company" → split on " at " │
│ source: jobs.ashbyhq.com/{co}/... → co │
│ source: jobs.lever.co/{co}/... → co │
│ source: boards.greenhouse.io/{co} → co │
│ source: {co}.myworkdayjobs.com/... → co │
└─────────────────────────────────────────────────────┘
│
▼
macOS popup: "Resume Builder — Starting for {Company}"
│
▼
┌────────────────────────────────────┐
│ Classification Council │
│ 3 × Haiku calls (parallel) │
│ │
│ each returns: │
│ ROLE_TYPE · EVIDENCE_FOR │
│ EVIDENCE_AGAINST · CONFIDENCE │
└───────────────┬────────────────────┘
│
┌────────┴─────────┐
│ All agree + │
│ ≥ 2 HIGH conf? │
└────────┬─────────┘
YES │ NO
│ └──► macOS dialog — recommendation + evidence
│ user confirms or picks different type
▼
Load archetype (if exists for role type)
else load experience files
│
▼
┌────────────────────────────────────┐
│ Headless Claude call │
│ (Sonnet — main model) │
│ │
│ outputs: KEYWORDS_A · KEYWORDS_B │
│ SKILLS LaTeX │
│ WORK_EXPERIENCE LaTeX │
│ EXTRAPOLATED table │
└───────────────┬────────────────────┘
│
▼
┌────────────────────────────────────┐
│ Keywords Council │
│ 3 × Haiku calls (parallel) │
│ │
│ each returns: │
│ APPROVED · MISSING_HARD │
│ MISSING_SOFT · CONFIDENCE │
└───────────────┬────────────────────┘
│
┌────────┴─────────┐
│ All approved + │
│ ≥ 2 HIGH conf? │
└────────┬─────────┘
YES │ NO
│ └──► macOS dialog — missing keywords
│ user adds or skips
▼
Compile PDF (pdflatex)
Warn in log if > 1 page
│
├──► ~/JobSearch/Resumes/{Company}_{Type}_{Date}.pdf
├──► ~/JobSearch/Applications/{Company}_{Type}_{Date}.md
├──► ~/JobSearch/tracker.csv (row appended)
└──► macOS popup: "Resume Ready — {Company}"
Council decision logic
Both councils (classification and keywords) use the same aggregation:
| All 3 agree? | HIGH votes | Confidence | Action |
|---|---|---|---|
| Yes | ≥ 2 | HIGH | Auto-proceed silently |
| Yes | 1 | MEDIUM | Show dialog with recommendation |
| Yes | 0 | LOW | Show dialog with recommendation |
| No | any | LOW | Show dialog with vote breakdown |
Role classification
| Label | Core signals |
|---|---|
| AIMLEngineer | MLOps · inference pipelines · model deployment · LLMOps · latency · scale |
| DataScientist | Statistical modeling · experiments · feature engineering · A/B testing · predictive accuracy |
| DataAnalyst | SQL · dashboards · KPIs · BI tools · business intelligence · stakeholder reporting |
| ProductAnalyst | Roadmaps · GTM · PDLC · cross-functional execution · sprint coordination |
Experience file selection
Three files always load. One optional 4th by JD domain:
Always loaded
├── itradenetwork.md
├── lia.md
└── digitas.md
Optional 4th (first match wins)
├── lexlead_ai.md ← llm, rag, agentic, mlops, modeling, genai, vector, embedding
├── teksystems.md ← nlp, text processing, document processing, ocr
└── galaara.md ← product, gtm, marketing, customer analytics, growth, retention
After 3+ applications of the same role type, run --compile to build an archetype. Future runs load the archetype instead, skipping raw experience files.
Vault structure
~/JobSearch/ ← outside ~/Documents to avoid macOS TCC restrictions
├── JDs/ ← drop job descriptions here (.md files)
├── Applications/ ← auto-generated per-application notes
├── Archetypes/ ← compiled role summaries (from --compile)
├── Resumes/ ← output PDFs
└── tracker.csv ← running log of all applications
Why not
~/Documents? macOS TCC restricts~/Documentsaccess for background processes. The launchd file watcher needs to read this directory; placing the vault at~/JobSearchavoids requiring Full Disk Access grants.
JD frontmatter — web clipper format (auto-parsed):
---
title: Support Operations Data Analyst at Harvey
source: https://jobs.ashbyhq.com/harvey/...
created: 2026-06-19
---
Or explicit:
---
company: Harvey
role: Support Operations Data Analyst
---
Repo structure
resume-builder/
├── skill.md # interactive orchestration logic (Claude Code)
├── roles/ # ATS writing rules per role type
│ ├── ai_ml_engineer.md
│ ├── data_scientist.md
│ ├── data_analyst.md
│ └── product_analyst_ops.md
├── experience/ # (gitignored) bullet libraries per employer
│ └── example_company.md # template — copy and fill in per employer
├── latex/
│ ├── template.tex # base template — fill in contact info and Education
│ ├── build.sh # compiles output.tex → PDF, moves to vault
│ └── output.tex # (gitignored) ephemeral per-run file
├── prompts/
│ ├── headless_prompt.md # main auto-mode prompt — structured block output
│ ├── council_classify_prompt.md # classification council (evidence-based CoT)
│ └── council_keywords_prompt.md # keyword completeness council
├── scripts/
│ ├── auto_resume.py # end-to-end auto pipeline (called by launchd)
│ ├── config.json # council settings + experience file routing
│ ├── run_auto.sh # launchd wrapper (sources shell env)
│ └── setup_watcher.sh # install / uninstall / test / reset watcher
└── vault_template/ # sanitized scaffold — copy to ~/JobSearch to start
├── JDs/example_jd.md
├── Applications/
├── Archetypes/
├── Resumes/
├── Tracker.md
└── tracker.csv # header only — rows appended by auto pipeline
What's gitignored:
experience/*.md— your actual bullet libraries (personal work history)latex/output.texand build artifacts — ephemeral per-run filesscripts/.processed_jds.jsonandauto_resume.log— local runtime state
Setup
1. Install LaTeX
brew install --cask mactex-no-gui
Requires pdflatex at /Library/TeX/texbin/pdflatex.
2. Set up the vault
cp -r vault_template ~/JobSearch
3. Create experience files
Copy experience/example_company.md for each employer. Name must match the key in config.json.
4. Fill in latex/template.tex
Add your contact info, LinkedIn, GitHub, and Education section. The pipeline writes only Skills and Work Experience — everything else comes from the template unchanged.
5. Install the file watcher
bash scripts/setup_watcher.sh install
Registers a launchd agent that watches ~/JobSearch/JDs/ and runs the full pipeline when a new .md file appears. Also fires every 5 minutes as a fallback.
Usage
# Drop a JD — pipeline runs automatically
cp "role-at-company.md" ~/JobSearch/JDs/
# Interactive via Claude Code
/resume-builder
/resume-builder ~/JobSearch/JDs/role.md
# Compile archetype (run after 3–5 applications per role type)
/resume-builder --compile DataScientist
/resume-builder --compile AIMLEngineer
# Watcher management
bash scripts/setup_watcher.sh status
bash scripts/setup_watcher.sh install
bash scripts/setup_watcher.sh uninstall
bash scripts/setup_watcher.sh test # run on latest JD immediately
bash scripts/setup_watcher.sh reset # clear all processed state
bash scripts/setup_watcher.sh reset "file.md"
# Manual run
python3 scripts/auto_resume.py "/path/to/JDs/role.md"
# Check logs
tail -f scripts/auto_resume.log
Token cost
| Step | Calls | Model | When |
|---|---|---|---|
| Classification council | 3 parallel | Haiku | Every run |
| Headless write | 1 | Sonnet | Every run |
| Keywords council | 3 parallel | Haiku | Every run |
| Total — no archetype | 7 | mixed | ~8,000–10,000 tokens |
| Total — archetype loaded | 7 | mixed | ~5,500–7,000 tokens |
Archetypes cut Sonnet input tokens by replacing raw experience files with a pre-distilled keyword + bullet profile.
config.json reference
{
"council_size": 3,
"council_model": "claude-haiku-4-5-20251001",
"council_keywords_timeout": 120,
"always_load": ["itradenetwork", "lia", "digitas"],
"optional_fourth": { "lexlead_ai": ["llm", "rag", "..."] },
"role_files": { "AIMLEngineer": "ai_ml_engineer.md", "...": "..." }
}
| Key | Effect |
|---|---|
council_size | Number of independent council calls per decision (default 3) |
council_model | Model for council calls — Haiku keeps cost low |
council_keywords_timeout | Per-call timeout in seconds for keywords council |