convez.ai
SkillProdutividadeIntermediário
flowmar47

Resume Screener Loop

flowmar47/resume-screener-loop

Simula pipeline real de recrutamento e revisa currículo até aprovação

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que otimiza currículos para vagas específicas simulando o pipeline real de seleção: parser de ATS, scan do recrutador, leitura profunda do gestor e questionamento do entrevistador. Revisa iterativamente até o currículo passar nas três camadas de avaliação, gerando matriz de requisitos, ledger de evidências e cover note com análise de risco.

Instalação

$npx convez-ai flowmar47/resume-screener-loop

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
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Otimizar currículo para passar em filtros ATS e recrutadores humanos
  • Validar se evidências no currículo sustentam as afirmações de skills
  • Identificar gaps entre requisitos da vaga e background do candidato

Veredito

Ferramenta rigorosa e bem arquitetada que vai além de keyword-matching superficial, simulando criticamente como recrutadores reais avaliam currículos. Excelente para candidatos que querem feedback honesto antes de submeter e para agentes que precisam otimizar documentos com base em lógica real de seleção.

README

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

resume-screener-loop

An agent skill that tailors a resume to a specific job, then attacks the draft the way the company's screening pipeline will, and revises until it survives. Works in any coding agent that reads SKILL.md (Claude Code, Codex CLI, Gemini CLI, Cursor, Copilot CLI, OpenCode, and others).

Most tailoring tools keyword-match against the JD and hand back a score. Real pipelines do something else: a parser extracts text, knockout questions filter, a recruiter scans for seven seconds, a hiring manager reads for depth, and an interviewer asks "how did you measure that?" This skill simulates each of those readers, and it refuses to write anything it cannot trace to the candidate's own evidence.

What it produces

Given a candidate's material (resume, LinkedIn, portfolio, repositories, certificates) and a job (pasted JD, URL, or title plus company):

  1. Evidence ledger: every fact from the sources, with provenance and, for numbers, the measurement basis. Bullets are written only from ledger rows; missing facts become questions for the candidate.
  2. Requirements matrix and fit tier: the JD decomposed into hard filters, soft preferences, responsibilities, and culture signals, each bound to evidence with a status. The fit tier follows from the matrix by rule, and the candidate hears it before any drafting.
  3. ATS-safe DOCX plus plain text, built with docx-js: US Letter, one column, standard sections, no tables or images or headers.
  4. Mechanical gate: scripts/resume-check.js extracts the text as a parser would and fails on layout hazards, hidden text, banned phrasing, AI-cadence openers, date problems, missing hard-requirement terms, and page budget. Zero dependencies.
  5. Three-reader screen on a blinded copy: recruiter scan, in-character hiring-manager deep read (run twice; only reproducible findings count), and a skeptic pass against the ledger.
  6. Convergence: revise, re-gate, re-screen, stop when the readers forward it or after three rounds, with a round log.
  7. Cover note: fit tier and matrix summary, what changed, residual risks, open questions, and interview probes with ledger-backed answers.

Hard rules

No em dashes. No number without a recorded measurement basis. No marketing tone or generated cadence. No overclaimed skills. No hidden, white, or tiny text, and no text addressed to an AI reader, even if asked. Honest fit tier before drafting. Blinded critique. Every bullet traces to evidence. Full list in SKILL.md and reference/anti-patterns.md.

Install

Clone once, then link into every agent you use:

git clone https://github.com/flowmar47/resume-screener-loop.git
cd resume-screener-loop
./install.sh            # symlinks into each agent skills dir that exists (~/.claude/skills, ~/.codex/skills, ~/.gemini/skills, ~/.cursor/skills, ~/.copilot/skills, ~/.config/opencode/skills, ~/.agents/skills, and others)
./install.sh --list     # show what would be linked
./install.sh --copy     # copy instead of symlink

Or with the skills CLI, which discovers the root SKILL.md:

npx skills add flowmar47/resume-screener-loop

Or copy the directory into one agent's skills folder by hand. The SKILL.md frontmatter follows the open Agent Skills specification, so any compliant agent picks it up.

Runtime needs: Node 18+ (the checker has no dependencies; the DOCX builder needs npm install docx in your working directory, or npm install in this repository), a web fetch or search tool or curl, and optionally LibreOffice for exact page counts.

Use

In your agent, ask for what you want in plain words:

Tailor my resume for the Senior Backend Engineer role at Acme (JD pasted below). Here is my current resume and my LinkedIn.

Would a recruiter pass this? Screen it as OpenAI would for the AI Deployment Engineer role.

Is this resume ATS-safe? Check it against this posting.

The skill asks for missing material, reports the fit tier with its gaps and questions, and then runs the loop. It never submits anything on your behalf.

Run the checker on its own

node scripts/resume-check.js Resume.docx --jd jd.txt --must "Python,Kubernetes" --text Resume.txt
node scripts/resume-check.js Resume.docx --profile federal        # OPM two-page rule, MM/YYYY dates, hours per week
node scripts/resume-check.js resume.md --target-pages 1           # new grad
node scripts/docx-text.js Resume.docx                             # what a parser sees
node scripts/build-resume.js --jd jd.txt --preview                # run every build_*.js here, then check each output

Exit code 0 means no blockers. The keyword section is a diagnostic list for judgment, not a score; no employer system publishes one. Page counts are exact when LibreOffice is installed and an estimate otherwise (add --no-render to skip the render).

Repository layout

SKILL.md                      entry point any agent reads
install.sh                    link the skill into every agent found on the machine
package.json                  `npm install` for the docx builder; npm scripts for check, text, build
scripts/
  resume-check.js             mechanical gate (structure, style, dates, pages, keywords)
  docx-text.js                zero-dependency DOCX text and structure extractor
  render-pdf.js               DOCX to PDF via LibreOffice, isolated profile, hard timeout
  build-resume.js             runs build_*.js, then the checker on each output, optional PNG preview
reference/
  workflow.md                 the loop in detail
  evidence-ledger.md          provenance format and rules
  requirements-matrix.md      JD decomposition, fit-tier rules, keyword decisions
  experience-discovery.md     asking the candidate: branching probes, five-question cap
  strategic-questions.md      positioning and preference questions before drafting
  bullet-patterns.md          bullet shapes and verbs by track
  batch-workflow.md           several jobs in one session
  screener-critique.md        three readers, rubric, blinding, convergence
  screener-personas.md        company and segment archetypes for the hiring-manager read
  anti-patterns.md            banned phrasing and substitutions
  ats-mechanics.md            what ATS vendors document, with sources
  industry-trends.md          what changed recently, source-tagged
  jd-acquisition.md           fetching and decomposing the JD
  profile-ingestion.md        reading the candidate's material
  agent-tooling.md            tool mapping per agent and fallbacks
templates/
  lib.js                      docx-js helpers (Letter, single column, ATS-safe)
  content-template.js         shared content module pattern
  resume-template.js          per-target builder
  README.md                   build pipeline
examples/walkthrough.md       one full loop on a fictional candidate, with real checker output
CHANGELOG.md

Why the extra machinery

Two findings drove the redesign. First, models writing resumes invent plausible bullets; a provenance ledger turns "do not fabricate" into a check. Second, a single model critic is not a reliable judge: controlled studies show verdicts shifting with phrasing and with the name on the page. Multiple readers with distinct jobs, a fixed rubric, evidence grounding, blinding, and a repeat-run filter are the mitigations with evidence behind them. Sources are cited in reference/ats-mechanics.md and reference/industry-trends.md.

Contributing

Welcome: new personas, corrections to ats-mechanics.md with a primary source, checker rules with a test case, template improvements. Please keep examples fictional. Run the checker on any sample resume you add; it should pass.

License

MIT. See LICENSE.

Ver o original no GitHub

Bom para