Language: English · 한국어

Good postings close fast. Catch them before they vanish.

This project has one goal: find the postings that truly fit your career, within 24 hours of going up. It doesn't flood you with keyword matches like a LinkedIn alert, and it doesn't submit applications for you like an auto-apply bot. Instead, it checks each posting's key requirements against the experience and results recorded in your CV, scores it out of 100, and keeps only what clears the bar. That rubric isn't a made-up rule set: it was backtested and calibrated on past real applications with known outcomes, including CV screens passed at Uber and Google. Once set up, it runs itself as a scheduled Claude task — every day it checks LinkedIn, Indeed, and your target companies' careers pages, then sends one email with the recommended postings and their scores.

v3 (2026-07). Rewritten after two months of daily production runs and the backtest above. What changed and why: CHANGELOG.md.

One full run, start to finish: the morning email arrives, and the tracking Sheet behind it holds a row per posting with the score, the reason behind that score, and any company the run added on its own.

All data shown is a demo: fictional persona, real public job postings, illustrative Status values.

A job-alert pipeline that finds new postings daily, scores how well each one fits your career, and keeps every result in one place.

1. It works like a headhunter retained for you alone. The way a homebuyer hires a buyer's agent, this tool watches the market strictly on your terms — the titles, locations, industries, and target companies you set. Every day it checks LinkedIn, Indeed, and your companies' careers pages, dedups every find against the full history, and sends one email with only the Top and Strong postings worth a closer look. Unlike a LinkedIn alert that maximizes what you see, or an auto-apply bot that maximizes how often you apply, it automates search and screening only — whether to apply stays your call.

2. It scores what your career can prove, not how many keywords match. From each JD it extracts the 4-8 requirements that actually separate candidates and checks them one by one against the experience and results in your CV. Competencies that fit everyone are excluded from scoring, and a must-have with no supporting evidence keeps a posting below the recommendation bar no matter how strong the rest looks. The rubric was calibrated through two months of daily production runs and a backtest of past real applications with known outcomes, including CV screens passed at Uber and Google. The score is a priority signal for where to apply, not a prediction of passing — and every posting comes with a Fit Reason laying out the verdict and the biggest gap.

All of that tuning is on the record: every change and its reason is in the changelog and the rubric.

At its core this isn't an application running on a server — it's a work order an AI agent carries out every day. There's no server to operate and no code to write. With your CV and application criteria at hand, first-time setup takes about 10 minutes.

Built and tested with Claude Cowork and these three connectors (a connector is Claude's official way of plugging into another service on your account):

• Chrome — browses LinkedIn, Indeed, and company careers pages
• Google Drive — reads and writes the Google Sheet that stores results
• Gmail — sends the morning results email

Any other AI agent works too, as long as it can browse the web, read/write a Google Sheet, send email, and run on a daily schedule. Below is the quick-start summary; the full walkthrough is setup.md.

1. Get the repo. git clone, or Code → Download ZIP on GitHub and unzip it anywhere.

2. Create the Google Sheet. Make two tabs named Jobs and Companies, and copy the Sheet ID from the address bar. Then enter your target companies straight into the Companies tab — it is the single source of truth for your watchlist (your-input/companies.example.md shows the format). The columns each tab needs are laid out in setup.md, Step 2.

3. Fill in your profile and settings. your-input/ ships two templates. Copy preferences.template.md to preferences.md and config.template.md to config.md, then replace the placeholders:

   • preferences.md — target titles, locations, industries, exclusions, eligibility, the email score threshold, and the work-hours timezone remote roles are judged against
   • config.md — the Sheet ID, your email, the alert schedule and alert timezone, the deep-scan weekday, and whether new companies get auto-added

   Your cv.md is generated, not copied: paste your existing CV's text into your-input/cv-original.md, have your agent convert it once into the scoring file cv.md, and review the result yourself. The your-input guide contains the conversion request, review checklist, and field explanations. cv.example.md shows a generated cv.md, and companies.example.md shows the Companies tab format (step 2); both are references, not files to copy. The templates contain only runtime values, so the daily run does not reread the guide. Paste the Sheet ID you copied into config.md. All personal files here are covered by .gitignore, so your real CV and settings never reach GitHub.

4. Hook the pipeline up to your agent. In Claude Code, two commands install the job-alert skill:

   /plugin marketplace add Journey-512/catch-before-jobs-vanish
   /plugin install catch-before-jobs-vanish@catch-before-jobs-vanish
   

   Then create a scheduled task telling it to run the job-alert skill, including your local repo path. On Cowork or any other agent, copy the prompt block from skills/job-alert/SKILL.md, change only the repo path in Step 0, and register it as a daily scheduled task (setup.md, Step 4).

5. Test it once. Trigger the task manually or wait for the first scheduled run, then check that postings landed in the Sheet and the Top/Strong email arrived. No matches is fine too — "No new matches. System alive." means the system is working. The full checklist is setup.md, Step 5.

Once set up, your part is simple: check the email each day, and update the Sheet only for the postings you actually applied to.

1. Check the email at your scheduled time. It carries only Top (85+) and Strong (70-84) postings. Headhunter postings are tagged [Headhunter], and newly added companies are flagged separately so you can veto them. For every posting, which requirements your career matches, the biggest gap, and which experience to lead with are recorded in the Fit Reason kept alongside it in the Sheet.
2. Open the Google Sheet when you want the full picture. Every posting that reached scoring gets a row, whether or not it was emailed. Postings below the cutoff and those marked Excluded (...) or Closed (date) stay recorded too, so you can always see why something didn't make the email.
3. Record your outcomes. When you apply, set the Status column to Applied; update it to Passed - CV, Rejected - CV, or Lost as results come in. If you applied through a referral, append (referral) to each status. The system never overwrites a status you typed yourself.
4. Review the numbers once a week. The deep-scan day email adds applications, CV-pass rate, and still-pending counts per score band. Cold and referral applications are tallied separately, and every figure carries its sample size. That's how you check how well the scores are guiding your application priorities — and adjust preferences or Skill calibration only when needed.

Even on a day with no matches, a "No new matches. System alive." email arrives. If no email comes at all, it doesn't mean zero matches — something broke in the scheduled task or the send. The full flow is described in How it works.

The daily scheduled task works through ten steps, grouped into four stages:

flowchart LR
  A["1 · Read your inputs<br/>CV · preferences · config<br/>+ Companies tab"] --> B["2 · Collect candidates<br/>search · normalize dates<br/>hard filters · dedup"] --> C["3 · Score fit<br/>grade JD requirements<br/>post-filter · liveness"] --> D["4 · Record and notify<br/>Sheet rows<br/>email · run report"]

It reads the three your-input/ files and the Sheet's Companies tab first, then collects candidates from LinkedIn, Indeed, and your companies' careers pages. Cheap checks come first — title, location, and language read straight off the listing, plus the match against past history — and only the surviving postings get their full JD opened and scored. Ordering the work by cost keeps search and analysis from being spent on postings that were never candidates.

Nothing is written to the Sheet or sent by email until collecting and deciding are done. Every write happens after all decisions are made, so a run that stops midway leaves outside data untouched. After each Sheet write, the rows are read back to confirm they landed exactly where intended — and a Status you typed yourself is never overwritten.

The shared logic is versioned in skills/job-alert/SKILL.md and the rubric. Korean translations are separate: pipeline and rubric. Your career and preferences are read from your-input/; your watchlist lives in the Sheet's Companies tab — a registry the pipeline reads every run and appends discoveries to, never overwriting what you typed — while the Jobs tab holds posting history and your application outcomes. So you personalize the pipeline by editing your CV, criteria, and the Companies tab, without touching the shared logic.

Safeguards added against problems found in production:

• Per-posting permalinks — instead of search-result URLs whose contents shift over time, it stores per-posting jobs/view/{id} links.
• Dedup against the full history — the LinkedIn job ID, the careers-page job ID, and the company plus normalized title are matched against every row ever recorded.
• JD text is data, never instructions — directives planted inside a posting to manipulate the AI are ignored; scoring uses only requirements and CV evidence.
• Re-check on high scores — postings above 80 are verified still open before the email goes out; a posting counts as closed only on an explicit closure notice or a 404.
• Failures are never silent — if a source couldn't be checked, the email names the missing coverage. Zero matches still sends a heartbeat, so a missing email always means the run itself broke.

The outcomes you record in the Sheet feed a separate feedback loop. When a score looks off, the Fit Reason shows why, and a Skill calibration line can be tightened; actual application results are tallied as CV-pass rate per score band, with cold and referral applications kept separate and sample sizes attached. The system measures and reports — it never changes the rules by itself. What to adjust is always your decision.

catch-before-jobs-vanish/
├── README.md · README.ko-KR.md      this page (English · Korean)
├── CHANGELOG.md                     what changed in each version, and why
├── setup.md                         the full walkthrough — first setup to testing and troubleshooting (EN + KO)
├── skills/
│   └── job-alert/
│       ├── SKILL.md                 the 10-step job-alert pipeline that runs every day
│       └── fit-scoring-rubric.md    runtime scoring method, weights, caps, and backtest grounds
├── .claude-plugin/                  config files for installing as a Claude Code plugin
├── docs/
│   ├── fit-scoring-rubric.ko-KR.md  Korean translation of the scoring rubric
│   └── skill.ko-KR.md               Korean guide to how the pipeline works
├── your-input/                      your personal input area — real inputs gitignored; templates and examples public
│   ├── README.md                    what goes in each file, and how
│   ├── companies.example.md         reference for filling the Companies tab — never copied locally
│   ├── config.template.md           copy target for config.md
│   ├── cv.example.md                what a generated cv.md looks like — reference, not copied
│   └── preferences.template.md      copy target for preferences.md
├── LICENSE                          MIT
└── .gitignore                       keeps your real personal files out of the repo

Can it be used for roles other than PM or PO?

The search-and-record pipeline applies to other roles too. The scoring rubric, however, was built and validated on PM/PO postings and real application outcomes — to use it for another role, review the rubric for that role, not just the target titles.

Will it catch every new posting?

No. One daily scan across a handful of sites can't guarantee full coverage, and sites sometimes block access or restructure their careers pages. The pipeline names the sources and search scope it couldn't check in the email, so a gap is never mistaken for success.

Does a high score mean I'll pass the CV screen?

No. The Fit Score is a priority for where to spend application time first, not a probability of passing. Real outcomes also ride on factors a JD and a CV can't show — applicant competition, referrals, how recruiters search. How well the score holds up is measured separately, from the outcomes you record.

Where are my CV and application records stored?

Your CV and settings live in your-input/ in your local repo and are never committed to GitHub. Your company watchlist, postings, and application records stay in your Google Sheet, and alerts in your Gmail account. The author runs no server and has no access to your data. How the AI agent and connected services process data, though, follows each service's own settings and policies.

Do I need Claude specifically?

No. Any agent that can browse the web, read and write a Google Sheet, send email, and run on a daily schedule can do the job. The current version was built and tested on Claude Cowork and Claude Code, so on another agent, verify the behavior yourself.

Does it cost anything to run?

The repo is MIT-licensed and there is no server cost. Your AI agent's and connected services' plans or usage limits still apply. A normal run targets roughly 25,000-30,000 tokens; the deep-scan day, which checks Strong and Soft companies up to the configured cap in each tier (8 per tier by default), uses more.

Is it okay to check LinkedIn and Indeed postings this way?

This repo explains collection principles only — it ships no procedures for bypassing site blocks or bulk harvesting. Use your own account, and check and follow each site's current terms and allowed scope. Responsibility for consequences such as account restrictions also stays with you.

I'm Journey MJ Lee, a product manager. I built this while preparing a move into global PM roles: instead of re-checking a stack of job sites every day, I wanted to receive only the postings my career could actually compete for, without missing any. It still runs every day, and it's public so that anyone preparing the same move can run it with their own CV and criteria.

Contact: LinkedIn · hemegi.lee@gmail.com

This is a personal project, not affiliated with or endorsed by LinkedIn, Indeed, Google, Anthropic, or any other company mentioned in this repository. Checking external services' terms of use and reviewing what the AI agent produces are the user's responsibility. The Fit Score is reference information for prioritizing applications — it guarantees neither CV-screen results nor job-search outcomes.

MIT. Fork it freely and adapt it to the way you run your own job search.