job-pipeline

End-to-end job application run over the Job Hunt Search tab — read every posting Claude judged fit without a browser, dry-run the forms by script, send the ones the user approves, record them in the f

SKILL.md

GitHub

job-pipeline — the whole pass, cheaply

Five commands, five short tables. Nothing is read through the Chrome extension; postings are fetched by script and forms are filled by script. Budget: one output line per job per stage.

0. Preconditions

  • Dashboard running (Job Hunt.bat or npm run dashboard), Search refreshed today (node scripts/funnel.mjs search --refresh if the cache is stale).
  • LinkedIn Easy Apply wanted → the automation profile is logged in (node scripts/apply.mjs --login, once).
  • Scratch directory for this run: $TEMP/claude/ (or the session scratchpad).

1. Screen (rules) — job-screen

node scripts/funnel.mjs screen --json $S/screen.json

Writes nothing: the dashboard applies the same rules on every read, so their hits are already out of the Search tab. Do not --commit — that makes today's rule hits permanent hides. Claude judges new cards about ten a minute: right after search --refresh, wait until the dashboard's queue is empty (claude:— on many lines means it is still working).

2. Read the postings — no browser

node scripts/jobs-fetch.mjs --screen $S/screen.json --buckets apply,check --out $S/fetched.json
node scripts/jobs-fetch.mjs --retry $S/fetched.json          # re-fetch the rows that errored (LinkedIn 429)

One line per job: idx | company — title | claude:<verdict>/<score> | !low-pay !applied-before | mode | lock | level/years | +stack | -neg | lang | pay:… | apply:<vendor> form:…. Claude's verdict comes first — it read the whole posting against src/verdict-criteria.md; the rest are regexes over the same text:

  • claude:skip → not a candidate. Report it with its reason; do not hide it (it waits in "Skipped by Claude", and a criteria change may bring it back).
  • claude:maybe, claude:…(old) or claude:— → needs-you: something the user should check, as in the "Check" stack.
  • claude:apply → a candidate, unless the line says otherwise:
    • !low-pay → needs-you (pay under the floor is the user's call);
    • !applied-before → needs-you (a role like one already applied to);
    • hybrid «…» / onsite / lock:… / lang:… that Claude did not mention → read the text (--text) and say what you found; never hide on a regex;
    • apply:portal / himalayas-account / indeed → needs-you (account needed).
  • CLOSED → hide: posting closed.

Hide only closed postings, with node scripts/funnel.mjs dismiss "<url>" --reason ".." --title ".." --company "..": a hide is permanent and teaches the screening rules (same role, same company).

3. Apply — job-apply

Split the candidates into two JSON lists (batch-ats.json, batch-linkedin.json) and dry-run them with apply-batch.mjs: it fills every form and sends nothing. It also refuses on its own what should not go out unasked — no current Claude apply, pay under the floor, a role like one applied to — and says why.

node scripts/apply-batch.mjs --fetched $S/batch-ats.json --out $S/results-ats.json [--answers $S/answers]
node scripts/apply-batch.mjs --fetched $S/batch-linkedin.json --out $S/results-linkedin.json

Show the user the dry-run lines (company, title, fields filled, missing answers) and send only what they approve: --submit --only <idx,…>. --force <idx> / --accept-low-pay <idx> only on their word. Applications go out in their name; nothing is sent unseen. Boards that e-mail a verification code (Greenhouse) run headed and need the user at the keyboard, so run those last and say so. The dashboard's "Apply" fills no forms: it only opens the posting.

Then clear the missing-answers rows: write $S/answers/<idx>.json, re-run with --only. A label that will recur goes into the answer bank once: node scripts/apply.mjs --remember "<label substring>" "<answer>".

Every applied row is already in the funnel (the script posts it). Do not add it again by hand.

4. Close the loop

  • A second posting of a role just applied to (same company, other board) needs nothing: the screening rules hide it ("already applied to this role").
  • node scripts/funnel.mjs stats and applied | head to confirm the count.

5. Report to the user

Four lists, one line each, nothing else:

  1. Applied — company, title, board, funnel id.
  2. Needs you — node scripts/funnel.mjs attempts (codes not typed, CAPTCHAs, account walls, missing answers — recorded by apply.mjs, also on the dashboard), plus hybrid-light roles and marketplaces (lemon.io, Jobgether, Proxify…) from the screen, with the URL.
  3. Claude said no — the claude:skip rows, one line each with Claude's reason (they stay in "Skipped by Claude").
  4. Failed — submit-failed rows with the error text and screenshot path.

Cost rules

  • Never get_page_text or screenshot a job page through the extension; jobs-fetch has the text.
  • Look at a screenshot only for a submit-failed, once per board.
  • Keep the whole run's model output under ~10 lines per job.