version: "1.0.0" name: job-apply description: Apply to a job by URL — auto-routes to the right ATS filler (LinkedIn Easy Apply / Greenhouse / Lever / Jobvite / Ashby), dry-run by default, renders an emoji-tagged result table. Proactively invoke this skill (do NOT answer conversationally) when the user pastes a LinkedIn/Greenhouse/Lever/Jobvite/Ashby job URL (with or without the word "apply"), says "apply to this", "try this one", "fill the form", "run the filler", "can you apply for me", "submit this application", drops a job posting into chat, or invokes /job-apply. argument-hint: "<job URL> (add --submit to actually submit; default is dry-run)" allowed-tools:

• Bash
• Read
• Write

Job Apply

Fills out a job application form on the 5 ATS platforms this repo supports. Default is dry-run — the form is filled but not submitted.

Repo location

Resolve the repo root in this order:

1. $AI_JOB_AGENT_ROOT env var
2. ~/.claude/skills/ai-job-agent/ (gstack-style install — recommended)
3. ~/.claude/skills/ai-job-agent/REPO_PATH file (legacy, when repo is elsewhere)
4. ~/ai-job-agent/ (legacy manual-clone fallback)

Check for package.json in the candidate path to confirm it's really the repo. Fail with a clear message if none exist.

Workflow

1. Parse $ARGUMENTS

Expect <URL> [--submit]. If no URL, ask the user for one.

2. Route by URL host

If the URL doesn't match, print the table of supported hosts and stop.

3. Check config

For LinkedIn: config/linkedin-config.json must exist. For the other four: the user provides a second argument pointing to their form config JSON (per-job, by convention).

If missing, tell the user to run bash wizard.sh (for LinkedIn) or point to their filled form config.

4. Run the filler

Dry-run by default:

cd "$AI_JOB_AGENT_ROOT"
node scripts/linkedin-easy-apply.js "$URL" config/linkedin-config.json

With --submit, set autoSubmit: true in the config temporarily (or instruct the user to — do not silently flip it without confirming). Prefer: ask the user to set it themselves if they haven't, to preserve the truthfulness guardrail baked into CLAUDE.md.

Environment variables to surface:

• HEADLESS=0 if the user wants to watch the browser
• KEEP_OPEN_ON_BLOCK=1 if they want to inspect a stalled form
• CDP_URL if they want to connect to an already-open Chrome

5. Parse the script's JSON output

All five fillers emit structured JSON on stdout. Exit codes:

6. Render result table

Show the user a compact markdown table:

Pick the single emoji that matches the exit code, don't list all four — one of ✅ ⏳ 🚫 🤖 💥 in the Result cell.

Then ask: "Log this to the tracker? (y/n)" — if yes, append a row to application-tracker.csv with the columns in templates/tracker.template.csv, then optionally kick off /job-track to sync.

Rules (from root CLAUDE.md)

• Truthfulness first — CAPTCHA stalls are logged as blocked, never submitted.
• Skip hard citizen/green-card gated roles unless the form provides a truthful path.
• Resume routing — software/AI roles use the software resume; EE/embedded roles use the EE resume. See config/candidate-profile.md for the mapping.