Champion Tracker

Champion Tracker

Track a maintained champion list for job changes. init records a baseline of each champion's current company + title; track re-scrapes and emits only the movers. You (the agent) score each mover's new company against the ICP rubric and write the verdict.

When to use

• "Tell me when our champions change companies."
• "Track these power users for job changes and score their new employers."
• A recurring cadence over a champion CSV (built from reviews, posts, or CRM exports).

How to run

Phase 1 — baseline (init)

Primary (Phantombuster LinkedIn Profile Scraper):

python3 ${SKILL_DIR}/scripts/pb_profiles.py \
  --agent-id "$PB_PROFILE_AGENT_ID" \
  --urls ${WORKSPACE}/champions.csv \
  --mode init \
  --output ${WORKSPACE}/baseline.json

champions.csv needs a linkedin_url column (a name column is used as a fallback label).

Phase 2 — detect changes (track)

python3 ${SKILL_DIR}/scripts/pb_profiles.py \
  --agent-id "$PB_PROFILE_AGENT_ID" \
  --urls ${WORKSPACE}/champions.csv \
  --mode track \
  --baseline ${WORKSPACE}/baseline.json \
  --output ${WORKSPACE}/movers.json

Emits {movers: [...], new_baseline: [...]}. A mover has old_company → new_company (and/or title). Movers with no usable new-company data get needs_review: true.

Degrade — no Phantombuster key (Playwright)

cd ${SKILL_DIR}/scripts && npm install && npx playwright install chromium
LI_AT="<your li_at cookie>" node ${SKILL_DIR}/scripts/pb_profiles_pw.mjs \
  --urls ${WORKSPACE}/champions.csv --output ${WORKSPACE}/snapshot.json

Run it once per cadence to produce a snapshot, then diff snapshots in-agent using the same old/new company+title comparison pb_profiles.py --mode track performs.

Score & emit (you, the agent)

For each mover, score the new company against the ICP rubric on a 0–4 scale, attach a verdict, and return a movers table. Movers flagged needs_review (no new-company data) score 0 and are surfaced for manual review. Persist new_baseline.json for the next run.

Outputs

• baseline.json — [{linkedin_url, name, company, title}] snapshot for diffing.
• movers.json — movers with old/new company+title + your 0–4 ICP score and verdict.

Credentials / env

• Required: none. The keyless degrade (pb_profiles_pw.mjs Playwright + an LI_AT cookie) is the fallback LinkedIn-profile source.
• Optional:
  • PHANTOMBUSTER_API_KEY (+ a LinkedIn cookie on the phantom) or APIFY_API_TOKEN — if set → managed LinkedIn profile scraper (higher volume/reliability — recommended). If not → keyless pb_profiles_pw.mjs Playwright with LI_AT.
  • LI_AT — LinkedIn li_at session cookie for the Playwright degrade.
  • PB_PROFILE_AGENT_ID — configured profile-scraper agent (or pass --agent-id).
  • SUPABASE_URL + SUPABASE_SERVICE_ROLE_KEY (or an Airtable key) — durable baseline store (degrades to a workspace baseline.json for a single cadence).
  • ANTHROPIC_API_KEY only if the LLM scoring isn't platform-provided.

Notes & edge cases

• Always run init before track — without a baseline there's nothing to diff.
• Proxy + throttle LinkedIn scrapes; randomize cadence to avoid blocks (the Playwright script jitters its per-profile wait).
• Dedup champions by normalized LinkedIn URL; rows lacking a profile URL are skipped.
• A move with no usable new-company data scores 0 and is flagged for manual review.
• Phantom result schemas vary by build; if a phantom uses non-standard field names, pass a downloaded result file via --results-json and re-map in-agent.