jobpipe

My own job search, run as a pipeline: find postings, screen them, have Claude judge what survives, tailor a resume, apply, and track what happened. One SQLite file holds all of it. Node 22, no dependencies, no build step.

poll  →  screen  →  judge  →  queue  →  tailor → build  →  apply  →  track
8,600    8,000 out   ~600 read   top 5     fact-checked       headless    CSV

The design decision that matters

Collection is free, so everything is collected. Judgment is expensive, so it sits downstream of a screen that costs nothing.

The hosted tool this replaced did it the other way round: its matching engine only ran on postings you had already spent your daily quota collecting, handed over newest-first, so you bought blind and ranked afterwards. Here a poll pulls thousands of postings in under a minute, deterministic rules cut ~93%, and a model reads a whole description only for the few hundred that are left.

Every knob is upstream of the model, and every rule-out is measurable: a rejection stores every rule that fired, so the cost of a rule is the number of postings it kills alone. That is how the rules get tuned — never by feel.

Stages

poll — src/sources/. Greenhouse, Ashby and Lever publish unauthenticated JSON for every open role; one GET per board. Workday has no such feed, so its adapter searches each site for target titles, drops what the screen would reject, and only then fetches descriptions. Remote boards (Himalayas, Remotive, RemoteOK, We Work Remotely) are many employers in one feed. Indeed comes through JobSpy, opt-in and low volume. Postings are never deleted: a closed role just stops having last_seen moved, which is how live and closed are told apart.

screen — src/screen.ts, src/place.ts. Named rules: level words, job family, specialisms, a stated range under the floor, clearance, office days, location. The bias is one-way — when in doubt, pass — because the judge is cheap and a wrongly discarded posting is never seen again. Each past bug is a pinned test (Roman numerals rescuing "Senior … II"; "London, UK" matching a state code; benefits boilerplate saying "remote" over an office location).

dedupe — src/dedupe.ts. The same job on the employer's board, a feed and Indeed three times over is one fingerprint (normalized company + title). Dedupe runs after the other rules, among copies that passed — run first, a London copy could beat the US-remote copy of the same job and the job was lost.

judge — src/judge.ts. claude -p, four at a time, one JSON object back: a 1–5 score with subscores, a remote-truth verdict quoting the posting (boards' "remote" labels have been wrong), and ghost risk from facts the store supplies — age, still listed, reposts. Unparseable replies are recorded as failures, never guessed at. A run that hits a usage limit resumes where it stopped.

tailor / build — the resume is edited against the posting (emphasis and wording only), rendered to PDF, and refused if it is not one page or is identical to the master.

autoapply — src/autoapply.ts. Unattended, within a narrow lane:

• only postings scored 4+, believed remote, low ghost risk;
• only through the employer's own Greenhouse/Lever/Ashby copy, found by probing likely slugs and matching the title, and that copy must pass the screen too;
• the tailored resume must pass a fact check: no number or proper noun the source documents don't already contain, header unchanged, no "managed spend";
• a fresh headless, profile-less Chrome per attempt via Playwright MCP, with hard stops for accounts, passwords, CAPTCHAs, ID or financial fields, and on-site or relocation requirements; demographic questions always left blank;
• counted only when it ends on the board's own host with a confirmation, and never retried — an attempt that died may have submitted.

track — applications keyed on company + role, synced byte-for-byte with a CSV tracker.

Install

npm install -g @toryr/jobpipe     # Node 22.18+; the command is `jobpipe`
jobpipe init               # starter profile files in ~/.jobpipe/profile

Polling and screening need nothing else. Judging, tailoring and autoapply call the claude CLI (Claude Code); Indeed needs uv.

Your files live in profile/: constraints.md (hard rule-outs), background.md (who you are, including honest gaps — those are what stop the judge calling everything a maybe), rules.json (payFloor), answers.md (what application forms ask), and config.json (where the tracker CSV, resumes and application folders go). An npm install keeps all of it, and the store, in ~/.jobpipe (JOBPIPE_HOME moves it); a git checkout keeps it beside the code, gitignored.

Running it

jobpipe poll [--scrape] [--bulk]
jobpipe screen [--all]
jobpipe judge --limit 40
jobpipe report            # out/YYYY-MM-DD.md, with "new since last report"
jobpipe queue
jobpipe tailor <id> && jobpipe build <id>
jobpipe autoapply --dry-run

From a checkout, run ./bin/jobpipe.ts directly — no build step; npm test runs the tests and npm run build makes the npm package. scripts/scheduled-run.sh is the daily run for a systemd user timer.

Registry

companies.json holds 74 boards, each probed live before it went in; verify re-checks them. The registry decides output quality more than the code does — once the adapters worked, which companies are in the file became the whole game. A bulk registry of thousands more is built locally by scripts/import-slugs.ts from a CC BY-NC slug list and is never committed.

Scope

LinkedIn is never polled or automated. Indeed is read through JobSpy at low volume for personal use. Everything else is a public feed an employer or board publishes to be read.