claude-apply

End-to-end job application automation piloted by Claude Code.

claude-apply is a toolkit for running your entire job search as a pipeline — scan ATS career pages, score offers with a small LLM call, and apply to them in a real browser, all orchestrated by Claude Code. The heavy lifting happens in deterministic JavaScript modules; Claude is only in charge of the loose, hard-to-script bits (reading the DOM, generating cover letters, deciding when to stop and ask).

What it does

• Scans Lever / Greenhouse / Ashby / Workable / Workday career pages via their public APIs (zero LLM cost).
• Filters offers by keyword (required_any, excluded_any in config/portals.yml).
• Scores individual offers with a stripped-down claude -p call (~$0.03/offer) using your CV.
• Applies to offers in a real Chrome window via Chrome DevTools Protocol (CDP):
  • classifies form fields using label + name patterns,
  • fills text/email/select inputs (with React-safe setters),
  • uploads your CV via Playwright CDP (bypasses all page-level file restrictions),
  • optionally generates a cover letter,
  • submits, detects the confirmation page, updates your tracker.
• Tracks applications in data/applications.md and data/apply-log.jsonl.
• Dashboards everything to a self-contained dashboard.html.

What it does not do

• No headless scraping of logged-in accounts. If an offer requires a login, claude-apply stops and asks you to sign in manually in the open Chrome tab.
• No captcha solving. Same — it stops and asks.
• No lying on forms. Cover letters and free-text answers are grounded strictly in your CV (config/cv.md).
• No stealth / anti-bot evasion. It runs in your own Chrome profile, as you.
• No scraping of non-API career pages. Only ATSes with a usable public API are auto-scanned (Lever, Greenhouse, Ashby, Workable, Workday). Custom career pages require a manual fallback or a new fetcher (see docs/extending.md).

Quickstart

git clone https://github.com/LeoLaborie/claude-apply.git
cd claude-apply
bash scripts/setup.sh

Then, from Claude Code, run /apply-onboard with the path to your CV PDF — it extracts your CV, builds config/cv.md + config/candidate-profile.yml, discovers ~30 target companies, and runs scripts/setup.sh non-interactively.

Or, manually:

1. Edit config/candidate-profile.yml with your identity, availability, preferences.
2. Edit config/cv.md with your CV (markdown).
3. Edit config/portals.yml with the companies to scan.
4. Reload your shell (source ~/.zshrc or ~/.bashrc).
5. Launch Chrome with CDP: chrome-apply.
6. Install the claude-in-chrome extension in that Chrome window.
7. From Claude Code: /scan, then /score <url>, then /apply <url>.

Architecture

          ┌──────────────┐
          │  portals.yml │
          └──────┬───────┘
                 │
                 ▼
    ┌─────────────────────────┐        ┌───────────────────┐
    │   src/scan (5 ATSes)    │◄──────►│ scan-history.tsv  │
    └──────────┬──────────────┘        └───────────────────┘
               ▼
        data/pipeline.md
               │
               ▼
    ┌─────────────────────────┐        ┌───────────────────┐
    │  src/score (claude -p)  │───────►│ evaluations.jsonl │
    └──────────┬──────────────┘        └───────────────────┘
               ▼
     (human / agent triage)
               │
               ▼
    ┌─────────────────────────┐        ┌───────────────────┐
    │  src/apply (Chrome CDP) │◄──────►│  applications.md  │
    │    + /apply command     │        │  apply-log.jsonl  │
    └──────────┬──────────────┘        └───────────────────┘
               ▼
         dashboard.html

Commands reference

Command	Purpose
node src/scan/index.mjs	Scan ATSes listed in portals.yml; append new offers to data/pipeline.md.
node src/score/index.mjs <url>	LLM-evaluate an offer; append to data/evaluations.jsonl.
node src/scan/explain.mjs	Trace which prefilter rule accepts or rejects a given title (npm run explain -- "<title>").
node src/apply/upload-file.mjs	CDP file upload helper (called by /apply).
node src/dashboard/build.mjs	Regenerate dashboard.html from data/ and reports/.
bash scripts/setup.sh	Interactive first-time setup (Chrome CDP profile + templates + rc).
bash scripts/check-no-pii.sh	Grep the tree for personal data patterns (CI gate).
npm test	Run the node test suite.
/apply-onboard [cv.pdf]	First-time setup — extract CV, build configs, discover companies, run scripts/setup.sh non-interactively.
/add-company <name or URL>	Discover and append a new company to config/portals.yml (no hand-editing YAML).
/scan	Claude Code slash command wrapping node src/scan/index.mjs.
/score <url>	Claude Code slash command wrapping node src/score/index.mjs.
/explain "<title>"	Trace which prefilter rule accepts or rejects a given title against your current config.
/tune-filter	Interactive calibration of title_filter against cached scan-history.tsv. No network calls.
/apply <url>	Claude Code orchestrator: open → classify → fill → upload → submit.
/dashboard	Rebuild dashboard.html from data/ and reports/.

Configuration

All user data lives under config/ and data/ — both are .gitignored. Templates live in templates/ and are copied by scripts/setup.sh on first run.

• config/candidate-profile.yml — identity, availability, CV paths, preferences.
• config/cv.md — plain markdown CV used by score and cover-letter generation.
• config/portals.yml — companies to scan + title filter.
• data/applications.md — tracker (Markdown table).
• data/pipeline.md — inbox of new offers from /scan.
• data/scan-history.tsv — dedup source of truth.
• data/evaluations.jsonl — output of /score.
• data/apply-log.jsonl — output of /apply.

For AI agents

Read CLAUDE.md first. It lays out the invariants (no PII in commits, no guessed values, stop on ambiguity), entry points, conventions, and where to find things.

See also docs/for-agents.md for typical workflows, patterns to follow, and anti-patterns to avoid.

Supported ATS

ATS	Scanner	Form fill	File upload	Notes
Lever	✅ auto	✅	✅ CDP	Dedup by URL; blocks re-submission for ~3 months.
Greenhouse	✅ auto	✅	✅ CDP	Splits first/last name; many optional subforms.
Ashby	✅ auto	✅	✅ CDP	_systemfield_* naming; custom questions are free text.
Workable	✅ auto	✅	✅ CDP	Public widget API; body is empty (no job description).
Workday	✅ auto	—	—	Scanner supports paginated public API; /apply support not yet implemented.
WTTJ	⚠️ partial	✅	✅ CDP	Aggregator — jumps to the real ATS in most cases.
Teamtailor, SmartRecruiters	❌	⚠️ partial	✅ CDP	Custom React form; standard fields classify correctly; add-ons may need help.
Custom career pages	❌	—	—	Manual fallback; PRs welcome (see docs/extending.md).

Contributing

See CONTRIBUTING.md. Short version: npm test, bash scripts/check-no-pii.sh, Conventional Commits, one logical change per PR.

License

MIT.

Credits

claude-apply was bootstrapped from the private workspace driving the author's own internship search, and its scan / score / dashboard layers are inspired by and initially derived from santifer/career-ops (MIT). The apply module, the CDP upload helper, the Claude Code commands, and the AI-agent documentation are original to this project.