GradPilot — New-Grad SWE Job Board

A personal job-hunt tool for the new-grad cycle. A company tier list doubles as a job board: automated fetchers pull new-grad SWE openings per company into results/{slug}.json, a single-file HTML page lets you browse them (per company or all at once), mark what you've applied to, and an AI skill simulates the automated resume-screening gate for each job against two resume profiles.

Quick start (after cloning)

Requires only Python 3 (stdlib — nothing to install).

git clone https://github.com/kypkk/GradPilot.git gradpilot && cd gradpilot

# 1) One-time setup
python3 scripts/discover_ats.py            # probe every company -> data/ats-map.json
                                           #   (gitignored, takes a few minutes)
cp resume/example.tex resume/backend-infra.tex   # then replace with your real resume either .tex or .pdf

# 2) Fetch jobs (A + B layers)
bash scripts/run_ab.sh                     # -> results/{slug}.json, log: refresh.log

# 3) Browse the board
python3 server.py                          # http://localhost:8000/swe-tier-list.html

Use server.py, not python3 -m http.server — the server also persists the "applied" toggle back into the JSON files.

The company tier list lives in the inline data object of swe-tier-list.html. After editing it, re-sync with python3 scripts/make_lists.py (and re-run discover_ats.py so new companies get fetched).

Steps 1–3 give you the script-only board. The AI parts — the LinkedIn/in-house browser dive and resume screening — run as the two skills below.

The two Claude skills

newgrad-job-refresh (.claude/skills/newgrad-job-refresh/)

The browser half of the refresh. Using a logged-in Chrome (Claude-in-Chrome), it deep-dives LinkedIn and the Tier 1/1.5/2 + in-house ("other") companies' own career sites — the places the script APIs can't reach — applies the same filters as the fetcher, dedups against what's already in results/{slug}.json by apply URL, and merges only genuinely new roles via scripts/add_jobs.py (stable -manual- ids that survive future fetch overwrites). It never runs the A+B scripts itself.

resume-screen (.claude/skills/resume-screen/)

Simulates the automated screening gate for a job. For each job × each resume profile (backend-infra, ai-mlops), it fetches the full job description from the apply URL, applies knockout filters first (work authorization / sponsorship, US location, graduation timing, degree, experience), then scores the content match 0–100 with missing keywords and honest fix suggestions. Results go to screenings/screenings.json (keyed by job_id, decoupled from the fetch pipeline) via scripts/save_screening.py, which also flips the job's screened flag in results/. The board overlays per-profile badges (pass / borderline / knockout) on every screened job and can sort all jobs by best screening score.

Using the skills (Claude Code / Cowork / other harnesses)

Both skills are standard Agent Skills (SKILL.md folders) under .claude/skills/, scoped to this project. The only hard requirement: the agent must run with this repo as its working directory, with bash + python3 available. The refresh skill additionally needs a controllable Chrome that's logged into LinkedIn; resume-screen needs your resume profiles in resume/ (.tex or .pdf) and a way to open each job's apply URL (browser tool or web fetch).

Claude Code (CLI):

cd gradpilot && claude

Project skills are discovered automatically. Just ask in natural language:

• refresh the job board / check LinkedIn for new grad roles → newgrad-job-refresh
• screen my resume against the stripe jobs / run resume screening for everything not screened yet → resume-screen

Claude Cowork (desktop): open this repo folder as the session's project folder — the skills are picked up the same way. Install the Claude-in-Chrome extension and keep a Chrome window logged into LinkedIn for the browser dive. For unattended runs, create two scheduled tasks whose prompts ask for the skill by name, e.g.:

• twice daily: "In , use the newgrad-job-refresh skill to do the browser deep-dive only (don't run fetch_jobs.py / prune_dead.py — schedule bash scripts/run_ab.sh separately, e.g. via cron/launchd). Add only roles whose apply_url isn't already in results/{slug}.json, then report what's new."
• offset by 30 min: "In , use the resume-screen skill to screen every job whose screened field is false in results/, against both resume profiles, and report the verdict breakdown."

Other harnesses (OpenClaw, or anything that supports Agent Skills): copy or symlink the two folders from .claude/skills/ into your harness's skills directory (e.g. ~/.openclaw/skills/), then run the agent with this repo as cwd and ask the same way. The skills are self-contained instructions — no Claude-specific tooling beyond a shell, a browser tool for the dive, and the scripts in this repo.

Changing the start-fetch date

Jobs posted before this date are dropped (except dateless Tier 1/1.5/2 roles). It's set in two places — keep them equal:

1. scripts/fetch_jobs.py — the MIN_POSTED constant at the top:

   MIN_POSTED = "YYYY-MM-DD"   # e.g. the start of your recruiting season

2. .claude/skills/newgrad-job-refresh/SKILL.md — the same date in the browser dive's filter step (step 3).

Resume profiles

Put your two profiles in resume/ as backend-infra and ai-mlops, each as .tex or .pdf (when both exist the skill reads the .tex). resume/example.tex is a starting skeleton. Real resumes are gitignored.

Scheduling

• A+B scripts: cron/launchd whatever you like — e.g. macOS launchd calling bash scripts/run_ab.sh twice a day. Needs outbound network; if the network is unreachable the fetcher aborts without writing (it never wipes the board).
• Browser dive + resume screening: schedule the two skills (e.g. Claude Cowork scheduled tasks) in an environment with a logged-in Chrome.