/find

FIND (live job search)

Step 1 — Read resume(s) AND search profile, then build the search plan

1A. Resume

Read every resume file in `resumes/` (skip the README and `search_profile.md`). Extract:

• **Target roles** — the user's most recent role title + adjacent titles (e.g. if they're a "Senior Data Scientist", also search "Staff Data Scientist", "Senior Product Data Scientist", "Lead Data Scientist"). Pull 3-6 titles total.
• **Top keywords / skills** — 10-15 from the skills section + repeated terms across bullets (technologies, methodologies, domain terms).
• **Default location** — the city/state from the resume's contact line.

If `resumes/` has no resume (only the README and/or `search_profile.md`), stop and tell the user to drop their resume in.

1B. Search profile (optional, but read every time)

Look for `resumes/search_profile.md`. If present, read the whole file as free-form prose. Apply it as overlays on the resume-inferred plan:

• **Locations:** if the search profile names locations, those replace the resume's default. (Example: profile says "Remote only" → drop the resume's city, search Remote only.)
• **Titles:** if the profile mentions specific titles or role types, narrow the title list to match. If it explicitly excludes some (e.g. "no people management"), drop manager-track titles.
• **Keywords:** add any domain interests from the profile (e.g. "climate tech", "LLM infra") to the keyword pool used for searches and scoring.
• **Salary floor:** if the profile gives a number, use it for scoring (jobs at or above floor get +1).
• **Deal-breakers:** if the profile lists exclusions (e.g. "no crypto"), drop matching jobs entirely from results, before scoring.
• **Company-size or stage preferences:** keep in mind for the score; small-company-only profile + a Fortune-500 result → drop or score low.

If the search profile contradicts the resume (e.g. resume is full of ML, profile says "I want to switch to PM"), the profile wins. The resume describes what the user *can* do; the profile describes what they *want* to do.

**No salary filter unless the search profile sets one.** Print the inferred plan as a one-line summary so the user can see it scrolling past, then proceed straight to Step 2 — no confirmation prompt.

---

Step 2 — Search (Chrome + WebSearch in parallel)

**LinkedIn via Chrome — for each target role, for each location:**

https://www.linkedin.com/jobs/search/?keywords=<role>&location=<location>&f_TPR=r604800&sortBy=DD

(`f_TPR=r604800` = last 7 days · `sortBy=DD` = date posted descending)

Read results. Extract per job: Company, Role, Location, Type (Remote/Hybrid/On-site), Salary (if shown), URL. Paginate up to 2 pages per search.

**WebSearch sweep (run in parallel with the Chrome searches):**

Build 3-5 queries combining target titles, top keywords, and locations inferred from the resume. Examples:

• `"<target title>" jobs "<location>" site:linkedin.com OR site:indeed.com 2026`
• `"<target title>" "<keyword>" "<keyword>" Remote jobs 2026`
• `"<adjacent title>" "<keyword>" "<location>" 2026`

Combine LinkedIn + WebSearch results into one pool.

**Rate limiting:** If LinkedIn throttles or shows a CAPTCHA mid-search, stop and tell the user. Work with results collected so far, or re-run later in smaller batches.

---

Step 2B — Indeed (only if enabled)

Check `resumes/search_profile.md` for `Indeed: yes` (in the Platforms section) or `Discovery: enabled` (in the Indeed section). If neither is present, **skip this entire step** — Indeed is off.

Indeed Lane A — active search

For each target role, for each location from the search plan:

https://www.indeed.com/jobs?q=<role>&l=<location>&fromage=7&sort=date

Also run each with `&l=Remote` (drop `radius`). `fromage=7` = posted in the last 7 days, `sort=date` = newest first. Paginate with `&start=10`, `&start=20` (up to 2 pages per search). Read each results page with `mcp__Claude_in_Chrome__get_page_text`. Extract Company, Role, Location, Salary (if shown), and the job URL (`/viewjob?jk=...` or `/jobs/view/...`).

**Hybrid coverage:** the local-area pass already returns hybrid roles because they are posted under the office city. Do not drop a candidate for being "Hybrid" — if the user's search profile allows Hybrid or On-site for that area, it is in-scope.

Indeed Lane B — recommendation harvest (best-effort)

After active search, glance at logged-in Indeed for any personalized module:

• `https://www.indeed.com/`
• Any visible "Jobs for you", "Because of your profile", "Profile match", "Based on your qualifications" module.

Extract Company, Role, Location, Salary (if shown), URL, the match label, and tag source as `Indeed Recommendation`. Do not click apply. **Expect little or nothing here** — Indeed is sunsetting the desktop Recommended Jobs feature, so this lane is a bonus, not a dependency. If it is empty, move on.

Indeed Lane C — Career Scout ingestion (mobile, user-fed)

Career Scout is Indeed's personalized recommender and it is **mobile-app only (iOS/Android)** — it cannot be driven from Chrome. Handle it as a user-in-the-loop source:

• Prompt the user near the start of `/jobs find`: "Career Scout is mobile-only. If you've run it recently, paste any roles it surfaced (links, or company + title) and I'll fold them in." If they have nothing, skip and continue (fail-soft, never block).
• When the user pastes Career Scout roles: tag each `Career Scout`. Resolve a working URL if they give only a name (via WebSearch). Run them through the same pipeline as every other source: dedup → company-site verification → score → add.
• Career Scout picks count as a personalized recommendation for scoring (+1 bonus), but still must pass the minimum score.

Indeed pacing + fail-soft (REQUIRED)

Indeed runs Cloudflare + DataDome, stricter than LinkedIn:

• Go slow: one Indeed URL at a time, never fire all titles/pages in a burst.
• **On ANY block** (CAPTCHA / "verify you are