Kol Engager Icp

KOL Engager ICP

Capture in-market leads from influencer audiences — qualified, not raw. KOL audiences are BROAD, so this skill is built around a strict ICP filter and a topic-relevance gate: engagers of a KOL's off-topic viral post are noise. The hard cost lever is one post per KOL.

Pipeline: select_kol_posts -> extract_engagers -> enrich_apollo -> score_icp -> dedup_history. Every paid step has a degrade path; the Apify steps are cost-gated.

When to use

• "Find leads from KOL/influencer audiences." / "Scrape engagers from influencer posts."
• After kol-discovery, or with a manual KOL list.

Quick start with example configs

Ship-ready (strict) ICP configs live in ${SKILL_DIR}/configs/ — copy one and tune rather than authoring from scratch (KOL audiences are broad, so these intentionally include hard exclude_titles + competitors gates):

• configs/icp.b2b-saas-revops.json — RevOps / GTM buyers at B2B SaaS.
• configs/icp.devtools-eng-leaders.json — engineering leaders buying developer tooling / platform / DevOps.

cp ${SKILL_DIR}/configs/icp.b2b-saas-revops.json ${WORKSPACE}/icp.json   # then edit
# default Apify actors are pre-wired (see "Known-good actors"); --actor is optional
python3 ${SKILL_DIR}/scripts/select_kol_posts.py --kol-urls "<kol>" --topic-keywords "rpa,automation" --estimate-only

Known-good actors

The Apify scripts ship sensible public-marketplace defaults, so --actor is optional. They are swappable — pass --actor <user~actor-name> to use any actor whose input takes the same fields. Defaults assume the harvestapi LinkedIn actor family (popular, public, no LinkedIn cookie required); verify current pricing on the actor's Apify Store page (the script also reads live pricePerUnitUsd from the actor when available).

Swap example: --actor apimaestro~linkedin-post-reactions-comments-engagements; output is normalized identically regardless of source.

How to run

1. Pick one topic-relevant post per KOL (select_kol_posts.py)

# Estimate Apify cost first (no spend)
python3 ${SKILL_DIR}/scripts/select_kol_posts.py \
  --kol-urls "https://linkedin.com/in/alice,https://linkedin.com/in/bob" \
  --topic-keywords "rpa,automation,workflow,orchestration" \
  --actor "harvestapi~linkedin-profile-posts" --estimate-only

# Confirm spend and pick (writes only posts that clear the topic-relevance gate)
python3 ${SKILL_DIR}/scripts/select_kol_posts.py \
  --kol-urls "https://linkedin.com/in/alice,https://linkedin.com/in/bob" \
  --topic-keywords "rpa,automation,workflow,orchestration" \
  --min-topic-hits 1 --days-back 30 \
  --actor "harvestapi~linkedin-profile-posts" --yes \
  --output ${WORKSPACE}/chosen_posts.json

Keyless degrade: gather a KOL's recent posts via web search (or hand them in) as a JSON list and pass --posts-file posts.json instead of --actor — the same topic gate + ranking runs.

The output's eligible:false rows are KOLs whose recent posts were all off-topic — skip them.

2. Extract reactors + commenters (extract_engagers.py)

# Cost estimate for the chosen posts
python3 ${SKILL_DIR}/scripts/extract_engagers.py \
  --post-urls "<post1>,<post2>" \
  --kol-source "https://linkedin.com/in/alice" --post-topic "rpa, automation" \
  --actor "harvestapi~linkedin-post-reactions" --estimate-only

# Confirm and run (Apify primary)
python3 ${SKILL_DIR}/scripts/extract_engagers.py \
  --post-urls "<post1>,<post2>" \
  --kol-source "https://linkedin.com/in/alice" --post-topic "rpa, automation" \
  --actor "harvestapi~linkedin-post-reactions" --yes \
  --output ${WORKSPACE}/engagers.json

Run extract_engagers.py once per KOL so each batch is tagged with the right --kol-source / --post-topic, then concatenate the JSON arrays.

Degrades (auto-selected when no Apify token, or force with --source):

• PhantomBuster: --source phantombuster --engagers-agent-id "$PB_ENGAGERS_AGENT_ID" (needs PHANTOMBUSTER_API_KEY + a LinkedIn cookie on the phantom).
• Keyless Playwright: --source playwright with LI_AT set. One-time setup: cd ${SKILL_DIR}/scripts && npm install && npx playwright install chromium. Scrapes both reactors and commenters (with comment text).

3. Enrich firmographics (enrich_apollo.py)

python3 ${SKILL_DIR}/scripts/enrich_apollo.py \
  --input ${WORKSPACE}/engagers.json --reveal-email \
  --output ${WORKSPACE}/enriched.json

Adds title/seniority/company/company_domain/company_size/industry (+ email with --reveal-email) via a two-phase Apollo call: (A) org-resolve a bare company name to its primary domain (cached per employer), then (B) People Match keyed by name + domain for a higher hit rate on broad KOL audiences. Without APOLLO_API_KEY it degrades to profile-only (seeds title from the LinkedIn headline) so scoring still runs. Optional DROPCONTACT_API_KEY gives an email fallback when Apollo reveals none. Enrichment costs credits — enrich the engagers you actually intend to score, not every reaction (use --limit).

4. Score ICP fit + intent (score_icp.py)

cp ${SKILL_DIR}/configs/icp.b2b-saas-revops.json ${WORKSPACE}/icp.json   # or icp.devtools-eng-leaders / scripts/icp.example.json; then edit
python3 ${SKILL_DIR}/scripts/score_icp.py \
  --input ${WORKSPACE}/enriched.json --icp ${WORKSPACE}/icp.json \
  --output ${WORKSPACE}/scored.json

Deterministic 0-100 score = ICP fit (0-60) + intent (0-25, comment>reaction x recency)

• topic (0-15), with a scoring_breakdown per lead and tiers A/B/C. Hard gates force tier C: competitor employer, an excluded title, or ICP fit below --min-icp-fit. By default only tier A/B are written (--keep-c to keep all). Tune --tier-a / --tier-b.

5. Cross-run dedup (dedup_history.py)

python3 ${SKILL_DIR}/scripts/dedup_history.py \
  --input ${WORKSPACE}/scored.json \
  --history ${WORKSPACE}/kol_leads_seen.csv \
  --output ${WORKSPACE}/new_leads.json

Drops leads already seen in prior runs (same in-market people engage many KOLs), records the survivors in the CSV ledger, and writes only the new ones. Add --supabase-table kol_leads (with SUPABASE_URL/SUPABASE_KEY) for a shared multi-machine ledger.

6. Final review (you, the agent)

Read new_leads.json (tier A/B, each tagged with kol_source + post_topic), sanity-check the top tier-A leads against the scoring_breakdown, and return the qualified table.

Outputs

new_leads.json — [{name, headline, title, seniority, company, company_domain, company_size, industry, email?, profile_url, engagement_type, comment_text?, post_url, kol_source, post_topic, icp_score, icp_tier, icp_gates, scoring_breakdown}], tier A/B, deduplicated by normalized profile_url across the whole run history.

Credentials / env

• Required: none. The keyless --source playwright + LI_AT path is the fallback engager source.
• Optional:
  • Engager source — if APIFY_API_TOKEN (or PHANTOMBUSTER_API_KEY + cookie) is set → managed actor (higher volume; cost-gated). If not → keyless Playwright (--source playwright + LI_AT).
  • APOLLO_API_KEY — two-phase enrichment (org-resolve + People Match); profile-only degrade without it.
  • MILLIONVERIFIER_API_KEY — enrich_apollo.py --verify deliverability; keyless syntax+MX check without it.
  • DROPCONTACT_API_KEY (email fallback); PB_ENGAGERS_AGENT_ID; SUPABASE_URL + SUPABASE_KEY (shared dedup ledger; degrades to the CSV ledger).

Notes & edge cases

• One post per KOL is the hard cost lever — select_kol_posts.py enforces it; never scrape every post.
• Topic-relevance gate (--min-topic-hits) is what keeps a KOL's off-topic viral audience out of your funnel. Raise it for noisier KOLs.
• Strict ICP tiering is the whole point here: KOL audiences are broad, so default to surfacing only tier A/B and keep --min-icp-fit honest.
• Cost gates: both Apify steps refuse to spend without --yes; use --estimate-only to preview cost first.
• Every lead is tagged with the KOL and post topic it came from — keep that for personalized outreach ("saw you engaged with Alice's RPA post").
• Proxy + throttle the LinkedIn paths; dedup runs both intra-run (by profile_url) and cross-run (the ledger).