Outbound Prospecting Engine

Outbound Prospecting Engine

The top-level playbook. It is mostly agent orchestration of the other outreach skills, plus one piece of unique deterministic glue — contact_cache.py — that dedups against a persistent cache so the engine is repeatable without re-spamming.

When to use

• "Set up outbound prospecting for [client]" / "build a lead gen engine targeting [ICP]" / "find and reach out to companies that need [solution]."
• When the user wants the whole pipeline, not a single signal or a single send.

The pipeline (agent routes each step to a sub-skill)

1. (agent) Define signal sources from the client's ICP/motion.
2. Detect signals (parallel) — route to the sub-skills:
   • hiring -> hiring-signal-outreach (detect_signal.py --signal hiring)
   • funding -> funding-signal-outreach (detect_signal.py --signal funding)
   • LinkedIn posts/commenters -> linkedin-post-research + linkedin-commenter-extractor
   • news/events/competitor engagers -> news-signal-outreach / web-automation
3. (agent) Qualify & score against ICP — multi-signal = highest; job-posting + funding = strongest intent.
4. Find contacts — find_contacts.py (Apollo by domain + ICP titles; verify) from a sub-skill, for the top companies.
5. Dedup — contact_cache.py against the persistent cache; contact only the NEW ones, skip already-contacted, add the new ones back.
6. (agent) Personalize per lead with email-drafting (the surfacing signal as "why now").
7. Launch — cold-email-outreach (lemlist/instantly) and/or linkedin-outreach (phantombuster/CSV).
8. (agent) Monitor & iterate — pull tool metrics, persist history, schedule weekly re-runs (platform Cron), report to channel.

How to run (the glue)

The unique script is the contact-cache dedup — run it between contact-finding and launch:

python3 ${SKILL_DIR}/scripts/contact_cache.py \
  --new ${WORKSPACE}/found_contacts.json \
  --cache ${WORKSPACE}/contact_cache.json \
  --update --output ${WORKSPACE}/to_contact.json

It matches on normalized email + LinkedIn URL, dedups within the batch too, and (with --update) appends the NEW contacts to the cache. Everything else is the sub-skills above, invoked by the agent at each checkpoint.

Outputs

• Per cycle: a qualified + deduped lead list with contacts and personalized sequences, a launched campaign in the chosen tool, and tracked metrics persisted for iteration; review tables to the channel at each checkpoint.

Credentials / env

• env.required: none. Signal detection is keyless, qualify/personalize is the agent, and the launch step routes to cold-email-outreach / linkedin-outreach (CSV export with no key).
• env.optional (all degrade): APOLLO_API_KEY — if set → contact-finding + funding/leadership data; else → keyless serp + pattern-guess. APIFY_API_TOKEN — if set → LinkedIn/event/ competitor signal scale; else → serp/web-automation. DROPCONTACT_API_KEY — email-finding fallback. MILLIONVERIFIER_API_KEY — if set → verify; else → local syntax/dedup (bounce risk). Send/launch — if a send key (LEMLIST_API_KEY / INSTANTLY_API_KEY, or PHANTOMBUSTER_API_KEY for LinkedIn) is set → launch; else → export a CSV to send manually (the keyless default). Store creds (SUPABASE_*, HUBSPOT_API_KEY) — if set → durable cache/metrics/CRM; else → workspace file-cache CSV ledger.

Notes & edge cases

• Human checkpoints are mandatory — after qualification, after personalization, and after each cycle — gate launches on these.
• The contact cache is what makes this repeatable — always dedup before contact-finding and before launch; write outcomes back.
• Multi-signal leads convert best — prioritize them; single social mentions are awareness-only.
• Run signal detection weekly; rotate proxy + throttle every scrape source; degrade any missing paid enrichment to the keyless backbone and flag reduced fidelity rather than stopping.