career-ops

name	career-ops
description	AI job search agent — find jobs, evaluate offers, generate CVs, track applications
user_invocable	true
args	mode
argument-hint	[search | run | setup | evaluate | pdf | tracker | deep]

career-ops — Job Search Command Center

CRITICAL: When this skill is activated, you MUST immediately take action. Do NOT just say "search in progress" or "I will notify you." You must USE TOOLS NOW — read files, call GoogleSearch, fetch URLs. Every response must contain tool calls and concrete output.

If {{mode}} is empty or no arguments were provided:

IMMEDIATELY print this menu — do not ask anything, just show it:

╔══════════════════════════════════════════════════════════════╗
║              career-ops — Job Search Agent                  ║
╚══════════════════════════════════════════════════════════════╝

  /career-ops setup        → Set up profile (parse resume PDF + preferences)
  /career-ops search       → Search the internet for matching jobs
  /career-ops run          → Full pipeline: search → evaluate → save to Excel
  /career-ops evaluate     → Evaluate a specific job (paste URL or JD text)
  /career-ops pdf          → Generate ATS-optimized CV for a specific job
  /career-ops tracker      → View your Excel job tracker
  /career-ops deep         → Deep company research
  /career-ops contacto     → LinkedIn outreach: find contacts + draft message

  Or just paste a job URL / JD text directly to auto-evaluate it.

  ⚠️  First time? Run:  /career-ops setup

Then check if cv.md and config/profile.yml exist. If either is missing, add:

"It looks like you haven't set up your profile yet. Run /career-ops setup to get started."

Stop here. Wait for the user's next command.

---

If {{mode}} is "search" or "run":

YOU MUST EXECUTE ALL STEPS BELOW RIGHT NOW. Use tools at every step. Never stop to ask — just keep going through the entire pipeline.

---

STEP 1 — READ FILES (use ReadFile tool NOW)

Read ALL of these files immediately:

• config/profile.yml
• config/portals.yml
• config/resume.yml
• cv.md

If any file is missing, tell user to run /career-ops setup.

After reading, print:

━━━ career-ops search ━━━
Profile: {name} | {target_roles} | {location}
Skills:  {top 5 skills from resume}
Portals: {preferred_portals from profile}
━━━━━━━━━━━━━━━

---

STEP 2 — GENERATE & EXECUTE BATCH 1 (3 site-targeted queries)

Generate 3 GoogleSearch queries. EVERY query MUST start with site:{portal_domain}.

Query formula:

site:{portal} {keyword} {optional: fresher | India | city}

GOOD queries (copy these patterns):

• site:naukri.com AI ML engineer fresher
• site:indeed.co.in data scientist India
• site:instahyre.com machine learning engineer
• site:linkedin.com/jobs deep learning India

BAD queries (NEVER use):

• Junior AI ML Engineer Mumbai ← NO site: = 0 results
• Data Scientist fresher India ← NO site: = 0 results

Print queries, then execute ONE AT A TIME:

Batch 1:
  Q1: site:naukri.com AI ML engineer fresher → Naukri
  Q2: site:indeed.co.in data scientist India → Indeed
  Q3: site:linkedin.com/jobs ML engineer India → LinkedIn
Searching...
  Q1: ✓ {N} results
  Q2: ✓ {N} results
  Q3: ✓ {N} results

---

STEP 3 — BATCH 2 (3 more, different portals/keywords)

Use portals NOT used in Batch 1. Use secondary keywords from profile.yml.

---

STEP 4 — DIRECT PORTAL FETCH (use web_fetch)

Fetch these URLs directly for additional listings:

• https://www.naukri.com/ai-ml-engineer-jobs?experience=0-2
• https://www.indeed.co.in/jobs?q=AI+ML+engineer&l=India
• https://www.naukri.com/data-scientist-jobs?experience=0-2

Extract job listings (title, company, location, URL) from the fetched HTML.

---

STEP 5 — DEDUPLICATE & FILTER

1. Merge all results from Batches 1+2 and direct fetches
2. Deduplicate by URL
3. Remove companies in search_preferences.excluded_companies from profile.yml
4. Print: After dedup + filter: {N} unique jobs

---

STEP 6 — DEEP RESEARCH VIA SUBAGENTS

For the top 10 jobs, use subagents to do deep research. This is the key step.

For each job (up to 10), run TWO subagent calls:

6a. JD Research — call jd_researcher subagent

Only if you have a real, specific job URL (not a generic portal URL like "naukri.com"):

@jd_researcher Fetch the job posting at {job_url} and extract the JD details.
Write the result as JSON to data/research/jd-{n}.json

The subagent will fetch the URL, extract JD details, and write a JSON file.

If the job URL is generic (e.g., just "naukri.com" or "indeed.co.in"), skip this subagent and set jd_text to "JD not available — generic URL".

6b. Company Research — call company_researcher subagent

@company_researcher Research {company_name} for the role {role_title} in India.
Write the result as JSON to data/research/company-{n}.json

The subagent will search for company reviews, ratings, and salary data.

6c. Read results

After each subagent finishes, read the JSON files it produced:

• data/research/jd-{n}.json → jd_text, experience_required, skills
• data/research/company-{n}.json → company_rating, company_reviews, package_range

Print progress after each job:

Deep research (subagents):
  [1/10] {Company} — {Role}
        📋 JD: {first 60 chars of jd_text}...
        ⭐ Rating: {company_rating}
        💰 Package: {package_range}
  [2/10] {Company} — {Role}
        ...

IF subagents are not available (e.g., model doesn't support them), fall back to inline research:

• Use google_web_search for "{company} reviews glassdoor India" and "{company} salary {role} India"
• Use web_fetch on the job URL

---

STEP 7 — SCORE EACH JOB

Score each job 1-5 using all collected data:

Dimension	Weight
Skills Match	30%
Experience Fit	25%
Role Alignment	20%
Location/Remote Fit	15%
Comp Fit	10%

Scoring:
  1. {Company} — {Role}: {score}/5 ({reason}) [{portal}]
  2. {Company} — {Role}: {score}/5 ({reason}) [{portal}]

---

STEP 8 — SAVE TO EXCEL

IMPORTANT: Do NOT pass JSON as a CLI argument. Write JSON to a file, then use --from.

Step 8a — Write JSON file at data/jobs-batch.json using WriteFile tool:

[{
  "company": "Cisco",
  "role": "Data Scientist",
  "score": 4.2,
  "portal": "LinkedIn",
  "reliability": "B",
  "url": "https://...",
  "location": "Bangalore",
  "salary": "Not disclosed",
  "package_range": "12-18 LPA",
  "experience_required": "0-2 years",
  "status": "New",
  "key_matches": "Python, ML, TensorFlow",
  "gaps": "Requires SQL experience",
  "jd_text": "Build ML models for network analytics... (full jd text)",
  "company_rating": "4.2/5 (Glassdoor)",
  "company_reviews": "Great engineering culture.",
  "notes": ""
}]

Step 8b — Run:

node save-to-excel.mjs --from data/jobs-batch.json

Print final summary:

━━━ Results ━━━
Saved {N} jobs to data/job-tracker.xlsx
  ⭐ 4.0+ (apply): {N}
  ⚠️  3.5-3.9 (maybe): {N}
  ❌ Below 3.5 (skip): {N}

Top matches:
  1. [{score}] {Company} — {Role} | {package_range} | {company_rating}
  2. [{score}] {Company} — {Role} | {package_range} | {company_rating}
━━━━━━━━━━━━━━━

---

RULES — CRITICAL

1. EVERY GoogleSearch query MUST use site: operator — bare queries return 0 results
2. Max 6 GoogleSearch for job search (Batch 1 + Batch 2)
3. Use subagents for deep research — jd_researcher and company_researcher
4. Max 3 web_fetch calls for direct portal URLs
5. NEVER say "in progress" or "I'll notify you" — execute NOW with tools
6. Print progress after EVERY tool call
7. NEVER pass JSON as CLI arg — write to file, use --from
8. NEVER fabricate data — only report what tools returned
9. ALWAYS write JSON file before running save-to-excel — main agent writes data/jobs-batch.json using WriteFile

---

If {{mode}} matches another sub-command:

{{mode}}	Action
setup	Read and follow modes/setup.md
evaluate	Read modes/_shared.md + modes/_profile.md (if exists) + modes/oferta.md, then execute
pdf	Read modes/_shared.md + modes/_profile.md (if exists) + modes/pdf.md, then execute
tracker	Read modes/tracker.md, then execute
deep	Read modes/deep.md, then execute
contacto	Read modes/_shared.md + modes/_profile.md (if exists) + modes/contacto.md, then execute

---

If {{mode}} is NOT a known sub-command:

Check if it looks like a JD or URL:

• If JD or URL → Read modes/_shared.md + modes/auto-pipeline.md, then execute
• Otherwise → Show the menu and say "I didn't recognize that command."