Kol Content Monitor

KOL Content Monitor

A weekly content-intelligence composite. It scrapes each KOL's recent LinkedIn + X posts, the agent labels each post with a 1-3 word topic, then a deterministic aggregator does the math (per-topic KOL/post/engagement counts, week-over-week spike detection, signal tagging). The digest narrative is the agent's.

When to use

• "What are the top voices in [our space] posting about this week?"
• "Track [list of founders/experts] and tell me what they're saying."
• "Find trending narratives I can contribute to."

Workflow

1. Intake / config

Gather the KOL list [{name, linkedin, x}], days_back (7; 30 first run), min_reactions (20; X floor = half), keywords. Persist as ${WORKSPACE}/kol-monitor.json.

2. Scrape LinkedIn (per KOL)

Reliable path: PhantomBuster (PHANTOMBUSTER_API_KEY + LinkedIn session cookie). Keyless degrade (one-off, best-effort) — the bundled Playwright helper:

node ${SKILL_DIR}/scripts/linkedin_posts.mjs \
  --profile "https://www.linkedin.com/in/foo" --name "Foo" --max 20

Set LINKEDIN_COOKIE (the li_at value) for better results. Keep posts with engagement >= min_reactions.

3. Scrape X (per handle)

Chain the sibling x-mention-tracker sub-skill with a from:<handle> query:

python3 ../x-mention-tracker/scripts/x_search.py \
  --query "from:foohandle" --since 2025-05-01 --max-posts 30 > ${WORKSPACE}/foo_tweets.json

Keep tweets with likeCount >= min_reactions / 2. Normalize LinkedIn + tweet records into one array of {topic, kol, engagement, comments, url, platform, text} — you (the agent) assign topic (1-3 words) per post.

4. Aggregate + tag signals (deterministic)

python3 ${SKILL_DIR}/scripts/aggregate_signals.py --input ${WORKSPACE}/posts.json \
  --history ${WORKSPACE}/kol_topics.csv --output ${WORKSPACE}/topics.json

Tags per topic: Convergence (≥3 KOLs), Spike (≥2× last week), Underdog (1 KOL), Controversy (high comment/reaction ratio). Reads + appends a per-topic weekly-count history CSV so Spike works across runs.

5. Synthesize digest (agent)

Read topics.json, pick best posts per topic, build the top-engagement table and emerging topics, write content-action recommendations (this week / next week), and render markdown to output_path (default kol-monitor-[DATE].md).

Outputs

A markdown digest (workspace file / Agent Teams attachment): tracked-KOL counts, trending topics with signal tags, best posts per topic + links, top-engagement table, emerging topics, recommended content actions. Post/topic history persisted for spike detection.

Credentials / env

• Required: none — both scrapers degrade keyless and synthesis is provider-swappable.
• Optional (paid, if-set/else):
  • APIFY_API_TOKEN — if set → reliable Apify X scraping via the x-mention-tracker sub-skill (cost-gated); else → that sub-skill's keyless Playwright (the default).
  • PHANTOMBUSTER_API_KEY (+ LINKEDIN_COOKIE) — if set → reliable/at-scale LinkedIn post scraping; else → the bundled keyless Playwright linkedin_posts.mjs (the default; LINKEDIN_COOKIE alone still improves it).
  • KOL-set discovery (SERP): DATAFORSEO_LOGIN / DATAFORSEO_PASSWORD (or SERPER_API_KEY) — if set → paid SERP to discover/expand the tracked-KOL list (e.g. "top voices in [space]"); else → the host agent's keyless web search (the default). Discovery feeds the KOL list only; it never gates monitoring.
  • ANTHROPIC_API_KEY / OPENAI_API_KEY — if set → optional script-side clustering/synthesis; else → the host agent does topic labelling + digest synthesis (the default; no LLM key needed).
  • SUPABASE_URL / SUPABASE_KEY — if set → Supabase cross-run spike history; else → workspace CSV (the default; first run emits no Spike tag).

Notes & edge cases

• LinkedIn and X are anti-bot heavy — prefer PhantomBuster / Apify; route Playwright fallbacks through a proxy and expect lower volume.
• X engagement runs lower than LinkedIn — hence the halved threshold.
• Spike detection needs persisted history; first run emits Convergence/Underdog/Controversy only (the aggregator reports spike_available:false).
• Schedule weekly to catch the week's peaks; dedup posts by URL across runs.