MLGalusha/job-sourcer

job-sourcer

This skill is the top of the application funnel. Its job is to turn "Mason wakes up and wants to know if any of the ~30 companies he cares about posted a new role in his lanes" into a one command answer. It does not apply to jobs. It does not run role-intake. It only surfaces candidates and queues them for Mason to triage.

This is a spec-only skill at initial build. The scraping step is stubbed until Mason picks a strategy (manual paste, RSS where available, headless browser, or a hosted scraper). The skill's contract and data model are the stable parts; the fetch layer is swappable.

---

When to invoke this skill

• Mason says "check for new jobs" or "what's new on the target list"
• A scheduled run (cron, launchd, or similar) fires this skill on a cadence Mason sets
• After target-companies.md is edited and Mason wants an immediate refresh

Do not run this skill if:

• profile/target-companies.md does not exist or is empty
• The fetch layer is not configured and Mason has not said "just do the manual pass"

---

Prerequisites

1. profile/target-companies.md — the list of companies to watch, with their job board URLs
2. profile/preferences.md — lane preferences, skip lanes, comp floor, location rules
3. data/sourced-jobs.jsonl — append-only queue (created on first run if missing)

---

Data model

data/sourced-jobs.jsonl is one JSON object per line, append-only. Each line:

{
  "id": "<ulid>",
  "t": "<ISO-8601>",
  "company": "<name>",
  "company_slug": "<slug>",
  "role_title": "<title>",
  "role_url": "<url>",
  "source": "<careers_page | greenhouse | lever | ashby | workatastartup | manual>",
  "first_seen_at": "<ISO-8601>",
  "lane_guess": "<founding | ai-engineer | fde | fullstack-ai | local-raleigh | off-lane | unknown>",
  "triage_status": "new",
  "notes": "<optional>"
}

triage_status is a closed vocabulary:

• new — surfaced, not yet reviewed
• queued — Mason wants to apply, needs role-intake
• skipped — Mason looked and passed
• applied — already has an applications/<app_id>/ entry

The sourcer only ever writes new entries. Status transitions happen via a separate triage step (not in this skill).

---

The pipeline

1. Read target list and preferences

Parse profile/target-companies.md into a structured list. Each company entry should surface:

• Name
• Slug
• Careers page URL(s)
• Lane tags (which of Mason's lanes this company could fit)
• Last scanned timestamp (if tracked)
• Notes

Read profile/preferences.md for:

• Skip lanes (do not surface roles in these lanes)
• Comp floor (filter out obvious below-floor postings where comp is visible)
• Location rules (remote-only, Raleigh-local, etc.)

2. Fetch (swappable)

For each company, fetch the current job listings. The fetch layer is abstracted:

• Manual mode: prompt Mason to paste the current listings for one company at a time. Slowest but zero-infra.
• RSS mode: for companies with public RSS feeds (Greenhouse and Lever both expose these), fetch and parse.
• HTML scrape mode: for companies without feeds, scrape the careers page with a headless browser or fetch plus a light HTML parser. Brittle but works.
• Hosted scraper mode: offload to a third-party jobs API. Expensive but reliable.

Initial implementation: manual mode only. The skill prompts Mason to paste the listings for each company that is due for a scan, and parses the paste into structured entries. Other modes are TODO.

3. Deduplicate against existing queue

For each candidate role, check if an entry with the same role_url already exists in data/sourced-jobs.jsonl. If so, skip. If not, it's new.

Also check data/applications.json for any app_id whose role_url matches. If so, mark the candidate as applied and do not append a new new entry.

4. Classify the lane

Use simple keyword heuristics on the role title and (if fetched) the first paragraph of the JD:

• "founding engineer" / "founding AI" → founding
• "AI engineer" / "applied AI" / "LLM engineer" → ai-engineer
• "forward deployed" / "solutions engineer" (at AI co) → fde
• "full stack" / "full-stack" / "product engineer" → fullstack-ai
• Raleigh / RTP / Durham location → local-raleigh
• Anything else → unknown

This is intentionally dumb. The real classification happens in role-intake when Mason triages.

5. Filter against preferences

Drop any candidate whose lane_guess is in Mason's skip lanes from preferences.md. Keep everything else, even if the lane is unknown.

6. Append to the queue

For each surviving candidate, append a line to data/sourced-jobs.jsonl. Use a ULID for id, ISO-8601 for timestamps.

7. Produce the daily digest

Write a short markdown digest to stdout (and optionally to applications/_digests/<date>.md):

# Sourced jobs — <date>

**New since last run: <n>**

## By lane
- founding: <n>
- ai-engineer: <n>
- fullstack-ai: <n>
- ...

## Top candidates (hand-picked)
1. **<Company> — <Role>** ([link](<url>)) — <one sentence why it looks good>
2. ...

## Full list
<company>:
  - <role title> — <url>
<company>:
  - ...

"Top candidates" are the three roles most likely to be worth Mason's time, picked by:

• Lane match to Mason's top preferences
• Company stage (Series A to Series C is the sweet spot)
• Freshness (posted in the last seven days if visible)

8. Update last-scanned timestamps

Optionally touch profile/target-companies.md with a "last_scanned" date per company, or write to a sibling data/scan-state.json if editing the profile file feels too heavy.

9. Report to Mason

Print the digest. Under 300 words. Mason should be able to skim it in 30 seconds and decide which ones get role-intake'd.

---

Rules and guardrails

• Never run role-intake from inside this skill. The sourcer's only job is to surface. Mason decides what gets intake'd.
• Never skip the dedup step. A re-run that re-adds the same role as new will corrupt Mason's trust in the queue.
• Never filter by comp unless comp is explicit in the listing. Most listings hide comp; inferring "probably below floor" from stage is lossy.
• Never scrape aggressively. Respect robots.txt. Use polite rate limits. If a company's site blocks the fetch, fall back to manual mode for that company and log a note.
• Never promote a candidate past new. Status transitions happen in triage, not sourcing.
• Voice rules apply to the digest. No em-dashes, hyphen compounds, tildes, or arrows in prose. Lists and tables are fine.
• The file data/sourced-jobs.jsonl is append-only. Never rewrite it. If a candidate needs updating (e.g., the role URL changed), append a new entry with a note referencing the old id.

---

Open questions (Mason to decide before first real run)

1. Fetch strategy per tier of company. Manual is fine for the top 5; RSS for the middle 20; HTML scrape for the long tail. Mason needs to decide the cutoffs.
2. Schedule. Daily, weekly, or event driven (only when Mason asks).
3. Digest delivery. Stdout, markdown file, email, or local notification.
4. Triage UI. How Mason moves entries from new to queued or skipped. Probably a small CLI or a web view. Until then, edit the JSONL by hand.

---

What this skill does NOT do

• Does not run role-intake. That's a separate decision.
• Does not write cover letters or resumes.
• Does not apply to jobs.
• Does not edit profile/target-companies.md except for optional last-scanned timestamps.
• Does not modify existing lines in data/sourced-jobs.jsonl. Append only.