Job Sourcing

Job Sourcing

Find, aggregate, dedupe, and legitimacy-check open roles against the user's targets. Turns "what's out there" into clean tracker rows ready for scoring, without ever submitting anything.

When to use

• "Find me jobs", "source roles this week", "any new postings at my Tier A companies?"
• Recurring sweep, run on a cadence to catch fresh postings (ATS pages update daily)
• After job-search-setup defines targets, or after the user adds a company/title

Inputs

• Reads: workspace/profile/target-roles.md (titles + variants, location, comp floor, deal-breakers), workspace/profile/target-companies.md (named accounts + company-ICP), workspace/pipeline/tracker.csv (to dedupe against and find the next free id)
• Needs from user: nothing required if profile/ exists. Optionally: a max number of roles to add, a recency window (e.g. "posted in last 14 days"), or a specific company/title to focus on. If profile/ is empty, stop and route to job-search-setup.

Workflow

1. Load targets. Pull title variants, location and remote rules, comp floor, and deal-breakers from target-roles.md. Pull the named A/B/C list and the company-ICP rubric from target-companies.md.

2. Search, preferring the source of truth. For each target, run web searches in this priority order (higher = fresher and real, not a stale aggregator repost):

   1. Company career pages. site:<company>.com careers, or the company's jobs URL directly.
   2. ATS boards. These host the real reqs. Search patterns:
      • Greenhouse: site:boards.greenhouse.io <company> or boards.greenhouse.io/<company>
      • Lever: site:jobs.lever.co <company>
      • Ashby: site:jobs.ashbyhq.com <company>
      • Workday: <company> workday careers (URLs contain myworkdayjobs.com)
   3. Aggregators. LinkedIn Jobs, Indeed, Wellfound/Otta. Use them to discover companies you didn't have listed and to catch roles not yet indexed elsewhere. Treat their postings as leads to verify against the company/ATS page, not as final truth. For discovery beyond the named list, search the company-ICP criteria + title (e.g. "Series B fintech" + "Product Marketing Manager" + "remote") to surface new companies, then add them.

3. Optional bulk scrape. For high-volume board sweeps, note the JobSpy integration (docs/integrations.md), which scrapes LinkedIn/Indeed/Glassdoor/ZipRecruiter in bulk to CSV. Purely optional. The skill works with web search alone. If used, feed its rows into the same dedupe + legitimacy steps below.

4. Dedupe. Collapse duplicates by company + normalized-role-title + url. The same req reposted on LinkedIn, Indeed, and the ATS is ONE row. Keep the highest-priority source URL (company/ATS over aggregator). Also dedupe against existing tracker.csv rows so you never re-add something already tracked. Normalize titles (strip "Sr.", "(Remote)", location suffixes, req numbers) before comparing.

5. Legitimacy / scam screen. Score each surviving role against the red-flag list below. Drop anything that trips a hard flag and annotate soft flags in notes.

   Hard flags (drop it):

   • Upfront payment, "equipment/onboarding fee", or a request for bank/SSN details to apply
   • Contact is a personal email (gmail/outlook) or "text this number on Telegram/WhatsApp" instead of a company domain or ATS apply flow
   • No verifiable company footprint (no real website, no LinkedIn company page, no funding/press)
   • Comp wildly above market for the title with a near-empty JD ("$65/hr, work from home, no experience")

   Soft flags (keep, but note):

   • Vague JD with buzzwords and no concrete responsibilities
   • Ghost job signals: reposted every 30 days for months, "evergreen" reqs, no named hiring manager, aggregator-only with no ATS/company link
   • Staffing-agency middleman obscuring the real employer
   • Salary "DOE" with no band in a state that requires disclosure

6. Assign ids and write rows. For each new real role, take the next stable id (J001, J002, and so on, never reused), and append a row to tracker.csv with status=found, date_found=<today>, source, url, comp_range if listed, location. Leave fit_score, priority, and next_action blank. job-fit-scoring fills those.

7. Save the raw JD. Create workspace/pipeline/applications/<id>-<company>-<role>.md from templates/application-record-template.md, and paste the full raw job description into the record (so later skills score/tailor against real text, not a memory of it). Slug the filename: lowercase, hyphens, no spaces.

8. Summarize. Report to chat: N new roles added (with ids), M duplicates collapsed, K dropped for legitimacy (with the reason), and any new companies discovered worth adding to target-companies.md.

Output

• Writes to:
  • workspace/pipeline/tracker.csv – new rows, status=found, stable ids
  • workspace/pipeline/applications/<id>-<company>-<role>.md – one per role, raw JD pasted in
• Format: a chat summary table (id | company | role | location | comp | source | legitimacy) plus counts (added / deduped / dropped). Dropped roles are listed with their flag so the user can override if you were wrong.

Rules & quality bar

• Never auto-apply and never auto-send. Sourcing ends at tracked rows. Application is a separate, human-approved step.
• Company/ATS pages beat aggregators. When the same role exists on both, keep the ATS/company URL. Aggregator copies go stale and inflate ghost-job noise.
• Every row is a real, open req. If you can't confirm the posting is live and the company is real, it doesn't get a row. It gets dropped with a reason.
• Respect deal-breakers at the source. If target-roles.md says "no on-site only" or "US-remote only", don't add roles that violate it. Note them as dropped rather than adding known-bad rows for job-fit-scoring to disqualify later. (Borderline cases: add and let scoring flag.)
• Stable ids, never reused. Even if a role is later rejected/withdrawn, its id is retired, not recycled.
• Don't fabricate comp or dates. If comp isn't listed, leave comp_range blank. Do not estimate into the cell (an ICP-based estimate belongs in notes, clearly marked).

Related skills

• Reads from: job-search-setup (targets), and appends alongside application-tracker's schema.
• Hands off to: job-fit-scoring (score and prioritize the new found rows). Run it immediately after a sourcing sweep. New companies discovered here feed back into job-search-setup / company-research.