lead-sourcing-web

Lead Sourcing — Web

Source leads from the open web by translating ICP trigger criteria into web search queries and scraper jobs. Output: accounts (and sometimes contacts) tagged with citation-grade trigger evidence (permalink + date). Where Apollo and LinkedIn surface databases, this skill surfaces the event.

Clarification Protocol

When you encounter any of the following, STOP and ask the user — do NOT make a silent assumption, and do NOT flatly refuse:

• Ambiguous or missing input
• A fork in approach (multiple valid frameworks, segments, scopes, or sources)
• A rule that would block progress
• An assumption that materially shapes the output
• Unverified data that a downstream step depends on

Present 2–4 concrete options with trade-offs and your recommendation.

Format:

Decision needed: <what's being decided> Options:

• A) <option> — trade-off
• B) <option> — trade-off
• C) <option> — trade-off My recommendation: <X> because <reason>. Confirm or pick a different option before I continue.

Refusing to proceed is NEVER the final answer. Every "block" must come with an override option. If the user picks an override, honor it on the first request — do not re-prompt the same block multiple times in one run.

Scope

This skill ONLY produces web-research-derived lead CSV records (sourced from job boards, press releases, public web). Do not expand into adjacent topics. If the user's request implies adjacent scope, trigger Clarification Protocol with options to (A) stay in scope, (B) note the adjacent need as a follow-up, (C) hand off to a different skill.

Verification Requirements (Derived Details Only)

The search action itself does NOT require web verification — web search and job-board scraping returns are first-party for sourcing. However, every derived person/company detail that this skill or downstream skills will state as fact (titles, seniorities, company size, technographics, funding) MUST be verified or marked [UNVERIFIED] before push.

• Tag derived details inline as [H/M/L][Sn].
• Source URL goes in the lead record's provenance field.
• Unverified derived details belong in ## Assumptions to Validate, NOT in the lead's main fields.

Sources are mandatory for derived claims. If you cannot cite a source, mark [UNVERIFIED].

When to Use

• "Find companies hiring Salesforce admins where the job post mentions migration"
• "Companies that issued an RFP for [topic] in last 90 days"
• "Find Directors of Engineering who appeared on technical podcasts"
• Triggers Apollo can't see — RFPs, press releases, job-post text, regulatory filings
• Target accounts are stealth-stage, regional, or niche-vertical
• Database-only sourcing produces empty firmographics

Legend — Confidence Score (Derived Fields)

• [H] High — verified by ≥2 independent primary sources within last 6 months
• [M] Medium — single primary source, OR multiple secondary sources agree
• [L] Low — inferred / single secondary source / older than 12 months

Quick Reference

Concept	Value
Modes	API (SerpAPI/CSE/Apify) / Manual (queries-only) / BYO (link list)
Citation rule	Every signal carries an evidence_url that resolves — no URL, no signal
Trigger taxonomy	stack-replacement / funding / leadership-hire / RFP / compliance / podcast-guest / OSS-activity
Half-lives	Press: 6mo · Job post: 3mo · Funding: 12mo · RFP: 90d
Trigger differentiator	Trigger IS the personalization hook — web-sourced records ship with hooks pre-constructed
Boolean precision	Site-restricted queries (site:linkedin.com/jobs, site:sam.gov) → primary sources
Output focus	Companies first, contacts rare — pair with data-enrichment / lead-sourcing-linkedin
Compliance	robots.txt respected; paywall/login-wall not scraped; aggressive same-domain scraping paced

Procedure

1. Confirm ICP grounding. Read ICP scorecard from icp-definition. For each trigger, identify canonical web source per taxonomy. Reference ${HERMES_SKILL_DIR}/references/ for trigger-source mapping.
2. Determine mode. Search-API + Apify → API mode. Else partial → mode-mix. Else → query-recipe-only output.
3. Translate triggers → search queries. Generate Boolean queries with site restrictions per trigger. Store as reproducible search_query_set. Examples: stack-replacement → site:linkedin.com/jobs "Salesforce admin" ("migrating from" OR "moving off"); RFP → site:sam.gov "vendor risk" (after:<date>).
4. Pre-flight: discover(). Run sample query (5–10 results) per trigger. Surface samples with provenance + cost. Default behavior: wait for explicit authorization. If the user requests skipping the sample step, trigger Clarification Protocol with options: (A) run sample then confirm (default safe), (B) skip sample and execute full batch within cost cap (flag run as confidence: low — sample-skipped), (C) raise cap then re-sample. Honor user override on first request; do not re-prompt. 4a. Search-criteria width gate (Clarification Protocol). If sample volume per trigger is far above or below the user's apparent intent, STOP and present options: (A) widen queries / loosen Boolean (warn: more aggregator noise + higher scrape cost), (B) narrow queries / add site restrictions (warn: may miss valid sources), (C) keep current. Recommend based on sample quality and trigger half-life. Honor pick on first response.
5. Execute search batch + extract. Run queries (API) or instruct user (manual). For each source URL: scrape via Apify; extract company, domain, sometimes contact. Domain resolution falls back to [unverified — needs check].
6. Normalize to Lead schema. Map fields per conventions. Stamp provenance_company: [verified: <source>:<url>]. Every signal's evidence_url is non-negotiable. Construct personalization_hook from the trigger source. Run ${HERMES_SKILL_DIR}/scripts/normalize_lead.py.
7. Dedup + push + summary. Dedup on company_domain for company-only records. Run ${HERMES_SKILL_DIR}/scripts/dedup_leads.py. Run ${HERMES_SKILL_DIR}/scripts/push_to_crm.py. Summary: queries, trigger breakdown, missing-contact report, recommended next skill (data-enrichment / lead-sourcing-linkedin).

Pitfalls

• Citing an unresolvable URL — re-check at push time; 404 = drop, not fudge
• Generic queries returning aggregators — restrict by site or use canonical-source whitelists
• Missing the trigger time window — 14-month-old "raised Series B" isn't a trigger
• Pushing accounts without contacts as outreach-ready — pair with data-enrichment next
• robots.txt violations — default is respect; if user explicitly requests overriding for a specific source, surface options (A) skip the source, (B) use the public cached version (e.g., archive.org), (C) override with confidence: low — robots-overridden flag and the user's stated basis recorded in provenance — do not re-prompt the same source twice in one run
• Confusing speculation for citation — "likely a buyer" is not a citation; the URL is

Verification

1. Every record's signals[].evidence_url resolves to a page describing the trigger as claimed
2. Every personalization hook's source_url matches the signal evidence URL
3. search_query_set is reproducible — same queries catch new entrants without rebuild
4. data-enrichment and/or lead-sourcing-linkedin recommended as follow-ups
5. [unverified] records routed to review queue with original URL preserved

Output Format

• Sections required: Lead CSV (per schema), Run Summary (queries, trigger breakdown, missing-contact report, recommended next skill), Assumptions to Validate
• Tables / fields: see references/csv-schema.md for the canonical Lead CSV column list and encoding
• Length target: depends on query scope (cost-capped per gtm.sourcing_run_record_cap); Run Summary ≤300 words by default
• File type: CSV (lead records) + markdown (run summary)
• Mandatory closing sections: Verification Notes, Assumptions to Validate, Next Step