Applywright
peselev/applywright
Pipeline agentic de candidatura automática, local-first, construído para Claude Code
Ver no GitHubUm agente de candidatura a emprego que roda localmente na sua máquina, integrado com Claude Code. Ele automatiza o pipeline completo: busca a descrição da vaga, escaneia injeção de prompt, avalia compatibilidade com seu CV, adapta o currículo, gera PDF e registra a candidatura em um rastreador — tudo sem enviar seus dados para a nuvem.
Instalação
npx convez-ai peselev/applywright1 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
- Candidatar-se a múltiplas vagas automaticamente sem expor dados pessoais
- Adaptar currículo e respostas de formulário por role com bullets pré-catalogados
- Rastrear e deduplicar candidaturas em CSV ou Notion
Veredito

Ferramenta robusta e bem-arquitetada para quem busca automação de candidaturas sem comprometer privacidade. Exige setup inicial (Python, pandoc, typst) e conhecimento técnico, mas o design local-first e o tratamento de prompt-injection mostram seriedade. Muito útil para quem aplica em volume.
README
A tradução para o português deste README ainda está na fila.
Applywright
An agentic job-application pipeline you run from your own machine. You paste a job URL (or a queue of them); the agent fetches the posting, scans it for prompt-injection, assesses fit against your CV and portfolio, tailors your resume, exports a clean PDF, and records the application in a tracker. You review and submit.
Built for Claude Code. The work runs locally. Your CV, bullets, and applications never leave your machine.
What it does
- Fetches the JD from a URL, with fallbacks (direct fetch, Jina reader, ATS-iframe detection, manual paste).
- Scans for prompt-injection in two layers: a mechanical script (invisible characters, known phrases, HTML-comment imperatives, AI-directed commands) and a semantic pass where the agent reads the JD for manipulation disguised as job requirements. Job postings are untrusted input; this treats them that way.
- Assesses fit against your CV, persona, and a library of tagged "master bullets," producing a scored verdict and picking the two bullets that best match the role.
- Tailors and exports your resume to PDF (Typst), swapping in the chosen bullets and a per-application UTM tag on your portfolio link.
- Tracks every application in a CSV (default, zero setup) or Notion (optional), and dedups so the same job is never filed twice.
- Drafts cover letters and application-form answers in your voice, on request, with a strict anti-AI-tell rule set.
Your data lives in one place: profile/
Everything personal (identity, CV, bullets, persona, learnings) lives in profile/, which is gitignored. The engine (skills, scripts, templates) carries none of your details; it reads them from profile/config.yaml at runtime. That means you can publish or fork the engine freely, and your information stays put.
The repo ships profile.example/ (a complete demo persona). Your real profile/ is created from it on setup.
Quick start
New machine (macOS, Linux, or Windows):
First install the toolchain: Claude Code, Python 3, pandoc, typst, and pipx.
- macOS:
brew install pandoc typst pipx(and install Claude Code + Python). - Windows (PowerShell):
winget install JohnMacFarlane.Pandocandwinget install Typst.Typst, then install pipx withpy -m pip install --user pipx(and install Claude Code + Python). Native Claude Code, no WSL.
Then clone, install the CLI, and bootstrap:
git clone <your-fork-url> applywright
cd applywright
pipx install . # installs the `applywright` command onto your PATH
applywright bootstrap # profile/ from the example, tracker init, output/ inbox/ temp/
If applywright is reported as not found after install, run pipx ensurepath,
open a new terminal, and retry. That command must be on your PATH for the agent
to use it; pipx ensurepath adds pipx's bin directory (~/.local/bin) to your
shell profile. Verify the whole toolchain with applywright doctor. Per-OS
details are in SETUP-WITH-AI.md.
On Windows this is the step most likely to need attention. If PowerShell reports
that applywright is not recognized after pipx install ., run pipx ensurepath,
close the terminal, and open a new one so the session picks up the updated PATH.
Then run applywright doctor from inside the repo folder to confirm the command
resolves and the toolchain is complete.
The fastest way to fill profile/ is the guided setup: run claude and say "set me up." The orientation skill walks you through config, CV, bullets, and persona, and saves progress so you can stop and resume. Or edit the files by hand:
profile/config.yaml: name, email, phone, portfolio URL, tracker modeprofile/cv.md: your resume, with{bullet_2}/{bullet_3}placeholders the agent fillsprofile/master-bullets.md: your tagged story bankprofile/persona.md: your positioning and case studies (or runrefresh-personaif you have a site)
Existing machine:
cd applywright
claude # opens Claude Code here; it reads CLAUDE.md automatically
Paste a job URL, or queue many in inbox/jobs.txt and say "process my inbox."
Daily use
Single job (auto by default): Paste a URL. The agent fetches, scans, assesses fit, then decides:
- Strong/Exceptional (≥6/10) → tailors the CV, exports the PDF, records the application as
To apply. - Weak/No fit (≤5/10) → records it as
Decided against applying. - No cover letter in auto mode. You review the folder and submit.
Want to weigh in? Say "manual" (or "pause" / "let me decide") with the URL. The agent then stops at the fit assessment so you can override the bullet picks or ask for a cover letter.
Bulk (always auto):
Drop URLs into inbox/jobs.txt (one per line), say "process my inbox," and the agent works top to bottom. You can keep appending URLs mid-run. Already-filed URLs are skipped (dedup); fetch failures are marked ❌ for manual retry. A roll-up prints at the end.
Post-processing (on request): cover letters and form answers. These never run automatically. Ask for them by company name or short-id once a job is filed.
- Cover letter: say "cover letter for {company}" (or the short-id). The agent drafts in your voice using
profile/cover-letter-field-notes.mdand the rules inskills/shared/writing-rules.md, exports a PDF, and saves both the draft and the PDF into the application folder. The greeting and sign-off are fixed strings the template depends on, so it leaves those alone. - Application-form answers: paste the question(s) from the form (for example "Why this company?" or a "describe a time when..." prompt). The agent answers from your persona and master bullets, in your voice, and saves them to the application folder so you can paste them back into the form.
What you'll see: files open for review.
After a step finishes, the agent opens the file it just produced so you can read it: the fit report, the cover-letter draft, the final PDF. Drafts are Markdown (.md). On macOS a .md file opens in whatever app is registered for it, which is plain TextEdit by default. Install a Markdown-aware editor so these render properly. VS Code is the common pick; set it as the default app for .md files. Final resumes and cover letters are PDFs and open in Preview.
Approval prompts: why it pauses
The engine runs multi-step Bash. Claude Code gates certain command shapes behind a yes/no prompt before they run. This is Claude Code's safety layer working as intended. The run has not failed; the agent is waiting for you. It will look like it stops mid-flow, sits on a prompt, then continues once you answer.
You'll see these most often:
cdplus output redirection in one compound command, flagged as "path resolution bypass."- Brace groups containing quotes, flagged as "expansion obfuscation." The JD-saving step writes frontmatter this way.
- First touch of a directory (for example creating an
output/{company}folder), which asks once per directory. - A command's first run (
applywright fetch,applywright export-pdf, and similar). Everything runs asapplywright ....
All of these are expected and safe to approve. At each prompt:
- Pick Yes to run that one command.
- Pick "Yes, and always allow..." or *"don't ask again for: applywright " to stop being asked for that pattern again.
Runs get quieter as you approve the patterns you trust. A bulk run of many jobs prompts a few times at the start, then settles into almost none. If you'd rather skip the prompts entirely, pre-allow the tools in Claude Code's settings.
Tracking: CSV (default) or Notion (optional)
Set tracker.mode in profile/config.yaml.
- csv: rows go to
output/applications.csvviaapplywright tracker. No setup. Columns:filed_at, short_id, company, role, url, source, stage, fit, comments, submission_date. Move applications through stages by editing thestagecolumn. - notion: rows go to a Notion database via the Notion MCP. Requires the MCP configured in Claude Code and two database IDs in
profile/config.yamlundertracker.notion.
Either way, the agent checks the tracker before filing and won't record a duplicate URL.
When scraping fails
Single job: paste the JD into inbox/jd.md (the MarkDownload browser extension preserves formatting well) and tell the agent. In bulk/auto runs there's no paste prompt; failed fetches are marked ❌ in jobs.txt for manual retry.
Layout
applywright/
├── CLAUDE.md # master instructions Claude reads every session
├── profile/ # YOUR data (gitignored): config, cv, bullets, persona, field notes
├── profile.example/ # demo persona; profile/ is created from it
├── output/ # filed applications + applications.csv (gitignored)
├── inbox/ # jobs.txt (bulk queue) + jd.md (paste fallback)
├── skills/ # workflow playbooks the agent loads as needed
├── src/applywright/ # the applywright CLI (installed via pipx install .)
├── pyproject.toml # package metadata for the applywright command
├── templates/ # Typst templates (CV, cover letter, document)
└── temp/ # scratch (gitignored)
Editing the pipeline
Behavior lives in CLAUDE.md and skills/*/SKILL.md. Edit those to change how it works; most tweaks need no code. The voice rules for written application materials are in skills/shared/writing-rules.md; tune them to your taste.
Privacy
profile/, output/, temp/ contents, .env, and the paste buffer are gitignored. Your CV, bullets, and applications are never tracked by git unless you change that.
Using Claude in the browser (no install)
You can run a lighter version of this without installing Claude Code by handing the repo and your profile to Claude in the browser (claude.ai). Useful for trying it before you install anything, or working from a machine that doesn't have Claude Code.
Setup, per chat:
- Upload your
profile/files (config.yaml,cv.md,master-bullets.md,persona.md, and the field notes), or paste their contents. - Give Claude the engine: upload a zip of the repo, or point Claude at this public repo so it can read
CLAUDE.mdandskills/. - Paste a job URL or the JD text.
The idea is the same (fetch, scan, assess fit, tailor, draft a cover letter or answers), with these differences from the Claude Code version:
- Nothing auto-opens. The browser can't open files on your machine. Claude returns the fit report, the tailored resume, and any cover letter inline, and you copy or download them.
- No
openstep, no local tracker writes. You update your CSV yourself. If you connect the Notion connector in the browser, Claude can record applications there. - PDF export runs in Claude's sandbox or on your machine. Claude can run the Typst export in its own sandbox and hand you a finished PDF to download, or return the tailored
cv.mdfor you to export locally withapplywright export-pdf. - No approval prompts. The browser sandbox runs without Claude Code's per-command yes/no gates, so the run doesn't pause the same way.
- More manual. Files move in and out by upload and download instead of living in one local folder. There's no bulk queue and no dedup against a local tracker.
Use this when you can't install Claude Code. For day-to-day volume, the Claude Code flow above stays faster.
A note on other agents
Applywright is Claude-Code-native: it relies on Claude Code's skill discovery, the open command (macOS), and a set of scripts that exist to keep file mutations off the shell. The materials (your profile/, the CSV, the templates) are agent-agnostic, so a port to another agent would reuse them and swap the engine. That port isn't here yet.
License
MIT. See LICENSE.