find-jobs

Search the web for new job postings that fit the user (at the level that matches their career stage, former employers always included). Use when the user asks to find jobs, look for openings, see what

SKILL.md

GitHub

/find-jobs

Usage: /find-jobs [role, company, or location] (or just ask in plain English).

Focus: $ARGUMENTS (blank = their preferences).

  1. Read profile/preferences.md, knowledge/field/target-employers.md (if it still says STATUS: not generated, do a quick web lookup of employers in the user's field and locations instead, and offer to build the file), jobs/tracker.csv, and jobs/leads.csv (every lead found before and what the user decided; create it with just the header row below if it's missing). A posting is a duplicate if it matches a tracker row or any lead (any status). Compare job IDs first: same company + same job ID is a duplicate; two different job IDs are separate openings, even with the same title and location. Then links: same link or apply link (ignore http(s)://, www., a trailing / and tracking bits like utm_..., gh_src, trk). Only when a job ID is missing on either side, same company + same title at the same location (check the posting text) counts as a duplicate; two blank job IDs never count as matching.
    • Learn from past decisions. Count not-for-me reasons and skipped leads in jobs/leads.csv. Never re-suggest a skipped or not-for-me posting. Tune the search: two or more too-senior → search one level lower and drop titles with Senior/Lead/II/III or "3+ years"; too-junior → the reverse; wrong-location → only the user's listed places (look at which places were rejected); not-my-field → leave out those role families; company → leave that company out; pay → skip postings that show pay below the rejected ones; other: ... → read the words and apply them. If a pattern contradicts profile/preferences.md (e.g. they keep rejecting a target role), tell the user in one line and offer to update preferences; don't change the file without a yes.
  2. Delegate the search to the job-scout agent with: target roles, locations, remote/hybrid/onsite preference, companies, the seniority that matches career_stage (e.g. "new grad / entry level" for students and new grads, "senior" for senior candidates), adjusted by the patterns above, the earliest start date, the kind of work (full-time, part-time or hourly with their availability, contract, academic), whether they need visa sponsorship, the list of already-tracked jobs and all past leads (company, title, location, link, status, reason), and a one-line summary of what the user turned down and why. Ask for up to 15 results.
  3. Rules for sources: for target employers, the employer's public job-board feed first (Greenhouse, Lever, Ashby, Workable, SmartRecruiters or Recruitee; addresses, date caveats, government boards and "Who is hiring?" threads are in references/sources.md; read-only, no login, no API keys). When the scout reports a board token for a company, you may record it in that company's Notes in knowledge/field/target-employers.md (e.g. "greenhouse: acme") so the next search is quick. Then company career sites (most reliable and where applications should go); for students and new grads, especially their university / early-career / new-grad pages. Then major job boards (Indeed and similar) and job boards specific to the user's field or region. LinkedIn job pages only if public; never scrape LinkedIn or Glassdoor. Never log in anywhere for the user. If the user is a student or new grad, remind them that Handshake (school login) lists campus-recruiting roles: they can paste those into /add-job.
    • For senior and executive roles, many openings are filled through executive search firms and are never posted. Say so once, and offer /network to draft a short note to two or three firms or recruiters who place people in the user's field (they're paid by the employer; any fee asked of the user is a scam signal; retained vs contingency and how to approach them: .claude/skills/network/references/recruiters-and-search-firms.md). Postings by agencies naming no employer get staffing agency in concerns.
    • Kind of search (from preferences, a current role in the master resume, or what the user said; ask once if unclear): employed and searching quietly, hourly or shift work, contract or freelance, or academic jobs each have a file in the playbook's "Special situations". Read only the matching one and apply its search notes (sources for each are in references/sources.md). For a user who is employed, never suggest anything public (no Open to Work frame, no public posts) and remind them once to search on their own devices.
    • Rank postings whose level and start timing match the user first; postings that need someone before the user's earliest start date go last, marked as such. Within the same fit, a warm path ranks higher: a contact there, or alumni counts in the employer's Notes in target-employers.md (.claude/skills/network/references/alumni.md); say it in why.
    • Connected company-data app (read only; rules and quota in knowledge/data-connectors.md): if the user asks, or the target list is thin, look for recently funded companies in their field and places, or ones backed by the same investors as a company they like. Those are employers to check, not postings: only real postings go in leads.csv. Pass what you found (with source and date) to job-scout as companies to search; it has no connector tools.
    • Always list any opening at a former employer (and, for students, a co-op/internship employer) from the master resume, even if the page didn't load, with the action "ask your former manager about it". Those are their best leads.
  4. Write every new result to jobs/leads.csv, one row each, appended with a targeted edit (re-read the file first; never rewrite other rows). Columns:
    id,found_date,company,role,location,remote,pay,posted,link,apply_link,job_id,source,fit_guess,why,concerns,status,reason,decided_date
    
    • id: L-<YYYYMMDD>-<NN> (today's date, numbered from 01 after any ids already used today). found_date: today.
    • remote: onsite, hybrid, remote, or blank. pay: the posted figure exactly as written (blank if none; never an estimate). posted: as the page says ("2026-10-03" or "3 days ago"), blank if not shown.
    • link: where it was found; apply_link: the employer's own application page if known. job_id: the requisition number if shown. source: e.g. greenhouse feed, company site, indeed.
    • fit_guess: A / B / C. why: why it fits, 12 words or fewer. concerns: short, e.g. stretch: asks 2 years, starts before you can, may not sponsor, needs a license you don't list, pay below your floor (top of the posted range under the salary floor in preferences; a note, not a skip), possible scam, staffing agency, only on job boards, couldn't open: check it's still open (the site blocked reading it; not proof it's closed), former employer: ask your manager. When judging a lead's concerns, use .claude/skills/evaluate-job/references/reading-postings.md (knockouts and "Real?" signals) so leads and add-job agree. A posting older than 30 days gets posted 30+ days ago only as a note, not a skip.
    • status: new; reason and decided_date blank.
    • Follow the CSV rules in the playbook ("Tracker"; quoting, plain punctuation, and no cell starting with =, +, -, @ or a tab). Then add one line to history/activity.md (e.g. Found 8 new job leads).
  5. Show the user a short ranked summary (best first), one line each (fictional example): A | Acme Corp | Operations Analyst | Austin TX | posted 3 days ago | why it fits in 6 words Mark anything that looks like a scam or a staffing-agency repost. Then the next step in one line: in Claude Code, "They're in your Job Inbox: type /job-inbox to save, skip, or pass on each one." In Codex (or if the panel isn't available), use the job-inbox skill in chat: "Tell me which to save, e.g. save 1, 3; skip 2." If the user asks how often to search: two or three short searches a week beat one long one, because early applicants get the closest look (daily while a campus, graduate-program or government intake window is open).
  6. When the user saves leads (a [Job Inbox] note from the panel, or they tell you), run /add-job with those links (it saves each posting with its deadline, job ID and apply link, scores it, preps packages for strong fits, and sets the lead's status to added).

Scheduled mode (the morning scan routine)

When the instructions say scheduled mode, first read .claude/skills/routines/references/scheduled-mode.md and follow it; it overrides any step above that needs the user.

  • Steps 1 to 4 as above with the saved preferences only: one job-scout call, up to 10 results, no follow-up searches. Skip the offers and questions (building target-employers.md, updating preferences from a pattern, the search-firm note, the kind of search when it's unclear); list the ones that came up.
  • Write only new rows in jobs/leads.csv (status = new, de-duplicated exactly as in step 1). Leave target-employers.md alone (board tokens wait for a run with the user), never run /add-job, and never touch jobs/tracker.csv.
  • Log: morning scan · <N> new leads (Job Inbox), or morning scan · nothing new. Summary: the count, the best one or two (company and role), any possible scam or former-employer opening, and "sort them in the Job Inbox".