job-finder

Files (8)

• SKILL.md
• job_sources.md
• company_targeter.md · prompts/
• job_board_searcher.md · prompts/
• link_validator.md · prompts/
• remote_specialist.md · prompts/
• salary_intel.md · prompts/
• scoring_rubric.md

SKILL.md (first lines)

# Job Finder

Single-shot pipeline: resume in, ranked job openings out. Designed for helping someone (a friend, family member, mentee, or yourself) look for work. Research and rank only, never auto-apply.

## What this is and isn't

This skill is a research helper. It scans public job boards and applicant-tracking system (ATS) endpoints, ranks postings against the candidate's resume, and produces a brief plus a tracker. It is not:

- A resume tailoring service
- A cover-letter writer
- An auto-apply bot
- A vetted-opportunity feed

Treat the output as a starting point. Verify every posting before applying. Listings on aggregators are often stale or republished.

## Trigger

- `/job-finder`, "find jobs for <name>", "job search for my friend", "look for jobs"
- Or any time the user hands over a resume PDF/image with a job-search framing

## Inputs

- **Resume path.** Default to most recent resume-shaped file in `~/Downloads/` (.pdf, .png, .jpg, .jpeg).
- **Optional overrides** (ask only if not obvious from the resume): preferred locations, salary floor, remote-only vs hybrid vs onsite, target titles to add or exclude.

## Configuration

Two placeholders to find-and-replace before you run this:

| Variable | Description | Default |
|---|---|---|
| `$JOB_SEARCH_ROOT` | Where to write deliverables | `~/Documents/Job Search/<Name>/` |
| `$AUTHOR_NAME` | Your name for the brief's "Prepared by" line | (your name) |

## Pipeline

### Phase 1: Parse resume

Read the file. PDF: `pypdf`. Image: use the `Read` tool (Claude vision handles PNG/JPG natively).

Extract a `candidate_profile.json` and write it to the run folder:

```json
{
  "legal_name": "...",
  "preferred_name": "...",
  "contact": {"email": "...", "phone": "...", "city": "...", "state": "..."},
  "current_role": {"title": "...", "company": "...", "since": "YYYY"},
  "target_titles": ["...", "..."],
  "years_experience": 0,
  "top_skills": ["...", "..."],
  "education": [{"degree": "...", "institution": "...", "year": "..."}],
  "salary_signal": null,
  "remote_preference": "remote|hybrid|onsite|unknown",
  "willing_to_relocate": true,
  "sectors": ["nonprofit", "higher-ed", "..."]
}
```

Heuristics:
- `target_titles` should reflect the resume's headline (the line under the name) plus close adjacents. Resist title-inflation. If the resume says "grant writer," don't auto-add "VP of Development."
- `years_experience` from the earliest role's start date to today.
- `willing_to_relocate` defaults to true unless the resume locks them to a region.
- `sectors` from employer types in work history (nonprofit, higher-ed, gov, startup, etc).

Sanity check: show the user the parsed name + top 3 target titles + location in one short message before fan-out. They confirm or correct. This prevents 4 agents wasting time on the wrong target.

Pseudocode:
```python
import pypdf, json
def parse_pdf(path):
    r = pypdf.PdfReader(path)
    return "\n".join(p.extract_text() or "" for p in r.pages)
# images: just Read the file via the harness; vision handles it
```

### Phase 2: Parallel agent fan-out

Spawn all four agents in a single message (multiple `Agent` tool calls in one block) using `subagent_type: "general-purpose"`. Each receives the candidate profile and its dedicated prompt file from `prompts/`.

| Agent | Prompt file | Mandate |
|---|---|---|
| `job-board-searcher` | `prompts/job_board_searcher.md` | Indeed, Glassdoor, Wellfound, ZipRecruiter, public LinkedIn search results. Current postings matching target titles + location radius. |
| `company-targeter` | `prompts/company_targeter.md` | Identify 15-25 mission-aligned employers and check their careers pages directly via Greenhouse/Lever/Ashby/Workable ATS endpoints. |
| `remote-specialist` | `prompts/remote_specialist.md` | RemoteOK, We Work Remotely, Working Nomads, Himalayas. Remote-only listings. |
| `salary-intel` | `prompts/salary_intel.md` | Market salary ranges for the target titles in the candidate's location, for context in the brief. |

Each returns JSON. Schema for the three search agents:

```json
[
  {
    "title": "...",
    "company": "...",
    "location": "...",
    "remote_status": "remote|hybrid|onsite",
    "salary_range": "$X - $Y" | null,
    "posted_date": "YYYY-MM-DD" | null,
    "url": "https://...",
    "source": "indeed|glassdoor|greenhouse|...",
    "raw_excerpt": "first ~200 chars of the posting",
    "fit_signals": ["matches grant-writing experience", "mission-aligned"]
  }
]
```

Salary-intel returns `{title, location, p25, p50, p75, source_notes}[]`.

### Phase 2.5: Link validation

Aggregator listings go stale fast. Indeed and ZipRecruiter routinely show jobs that have already been filled, expired, or were republished from years ago. To prevent dead links in the brief, run a fact-check pass after fan-out completes.

Merge results from the three search agents. Dedupe by `(company.lower(), normalize(title))`, keep the entry with the most complete fields. Score a quick preliminary fit (0-100) using `skills_match * 0.5 + experience_match * 0.5` only, fast heuristic. Take the top 50.

Spawn the `link-validator` agent (`prompts/link_validator.md`, `subagent_type: "general-purpose"`) on those 50. The validator:

- WebFetches each URL.
- Classifies the response:
  - `live`: posting page loads, title and company match the listing

From jasonbhorne/claude-code-skills, MIT.

More from jasonbhorne/claude-code-skills