Icp Prospect List Builder

ICP Prospect List Builder

Overview

How to build a prioritized, source-cited prospect dataset — hundreds of companies with decision-maker contacts — entirely from public data. The output is two cross-linked tables (companies + contacts) where every company is categorized, scored for ideal-customer-profile (ICP) fit, and every claim is attributed and confidence-tagged.

Use this skill when you need to:

• Find "companies like our existing customers" and turn them into a ranked target list
• Enumerate a market segment (e.g. "all website/app builders") across many sources
• Score a raw company list for fit against a specific product's ICP
• Attach the right contacts to each company (execs for small cos, functional leaders for large)
• Produce a CRM-importable dataset, not a one-off report

Developed and validated July 2026 building a 621-company / 900+-contact GTM dataset for a custom-domain-onboarding product (Connect Domain, an Entri alternative), spanning disclosed-customer research, 10 accelerator portfolios + Y Combinator, a Minnesota regional ecosystem, and a Product Hunt API sweep — run with the parallel-agent-fanout skill.

Ethics (non-negotiable): public sources only; professional/business contacts in their professional capacity only (no personal/home data); never fabricate — a blank field beats a guess; every row carries a source_url + confidence.

Step 1 — Define the ICP in searchable terms

A fit definition you can't query is useless. Convert the ICP into (a) a one-sentence buyer description, (b) priority sub-segments, and (c) a keyword lexicon that appears on a matching company's site.

Worked example (custom-domain onboarding product):

• Buyer: multi-tenant SaaS platforms whose end-users connect their own custom domain during onboarding ("bring-your-own-domain").
• Sub-segments: website/site builders · AI website/app builders · e-commerce builders · newsletter/email + deliverability · link-in-bio/landing/funnel · publishing/CMS · course/community/membership · forms/booking/portfolio · white-label/agency/PaaS.
• Lexicon (recognize fit on-page): custom domain, connect your domain, bring your own domain, add a custom domain, custom sending domain, white-label domain, vanity domain, map your domain, SSL for SaaS.
• Anti-signals (exclude): hardware, robotics, biotech, deep tech, pure data/AI infra, consumer mobile apps, logistics, single-tenant internal tools.

The lexicon is what every downstream search and every ICP-fit judgment keys on.

Step 2 — Discover companies (sources ranked by signal)

Sweep sources in roughly this order; higher tiers give cleaner ICP density.

Tier 1 — highest signal

• Source-code / technographic search — the single best move. Search page HTML for a competitor's embed string to get their live customer list: PublicWWW (publicwww.com), NerdyData, BuiltWith relationships. E.g. searching for a rival's widget script returns every platform that embeds it — literally your ICP. (These bulk-index tools are usually paywalled; a paid seat is the highest-ROI data buy. Log the cap when gated.)
• Competitor customer bases — scrape case studies / logo walls / testimonials from adjacent vendors. Their customers are your ICP. Also mine any relevant open adopter list (protocol adopter pages, "customers of X" pages).

Tier 2 — high volume, well-categorized

• Software directories — G2, Capterra, GetApp, TrustRadius. Category pages map 1:1 to sub-segments; pull the vendor list, not reviewers. (G2 often 403s bots; fall back to category listicles / search snippets.)
• Product Hunt (API) — best for newly launched tools. Use the GraphQL API, not scraping. See scripts/producthunt_pull.py. Query the root posts(topic:) field (the Topic type has no posts connection), order by VOTES, drop nested topics from the node (it inflates query complexity). Rate limit ~6250 complexity / 15 min — self-throttle on the x-rate-limit-remaining header. Note: the website field is a tracked producthunt.com/r/... redirect that 403s to bots, so resolve real domains later (during the contact pull) rather than from the API.
• AI-tool directories — There's An AI For That, Futurepedia, Toolify. Best for the AI-builder wave the other sources lag on.

Tier 3 — supplementary

• Crunchbase industry filters · AlternativeTo (enter a known ICP tool → get its cluster) · app marketplaces / partner directories · Indie Hackers / BetaList / "Show HN" · GitHub awesome-lists.
• Accelerator portfolios — filter each portfolio to ICP only. Reality check: ICP density is low in most accelerators (they fund biotech/hardware/fintech); a broad-thesis program (e.g. Y Combinator) is the exception. Filtering the full portfolio beats sampling it — but directories are often JS/Crunchbase-gated, so log coverage honestly.

Step 3 — The data model (two grains, cross-linked)

Maximum utility comes from two tables joined by a stable company_id:

companies.csv — one row per organization:

company_id, dataset, segment, organization, domain, description, category,
tags, icp_fit, icp_rationale, size_or_stage, location, num_contacts,
source_url, confidence

contacts.csv — one row per person (carries company_id + joined category/icp_fit so people are filterable by company attributes):

company_id, dataset, segment, organization, contact_name, title,
linkedin_url, email, email_status, source_url, confidence

dataset = the source pass (e.g. entri-customers, accelerator-icp, producthunt-icp) so pipeline slices stay separable. email_status ∈ public | inferred-pattern | none.

Step 4 — Categorize + ICP-score

Assign every company exactly one category from a controlled vocabulary (normalize messy free-text sectors into it) plus free multi-value tags. Rule: pick the real vertical, not a surface trait — an AI drug-discovery startup is Healthcare & Life Sciences, not AI/ML.

Then the prioritization layer — icp_fit (High / Medium / Low / n-a) + a one-line icp_rationale:

• High — the product clearly exhibits the ICP behavior (e.g. end-users attach their own domain). Existing customers of a direct competitor = High.
• Medium — plausibly could, but doesn't obviously center on it.
• Low — no ICP surface.
• n-a — not a buyer (investors/accelerators; relevant as channel only).

A deterministic keyword pass gives a fast backbone; a review pass (human or agent) corrects it and writes the rationale. This is what converts a directory into a ranked target list.

Step 5 — Dedupe

• Prefer domain as the dedup key (normalize: strip scheme, www., path, query). Fall back to a normalized company name when no domain (e.g. Product Hunt launches) — strip version/suffix tokens (2.0, v3, by <maker>, trailing AI/App) so relaunches don't read as net-new.
• Always dedupe against the existing master before adding a new pass; tag each candidate net-new vs already-in-dataset. Report the split — it is the honest signal of how much a source actually added.

Step 6 — Pull contacts (routed by company size)

• Small (<50): the executive team — founders / CEO / CTO / C-suite.
• Mid / large: the operations, financial, product, and technical leaders (CPO/VP Product, CTO/VP Eng, COO/VP Ops, CFO) — and for very large orgs, only the leaders of the relevant org (e.g. the websites/domains product line), not the global CEO.
• Per contact: name, title, linkedin_url (only if confidently the right person), email (public if published; a clear corporate pattern may be filled and marked inferred-pattern; else blank), source_url, confidence.
• Do not pad. A tiny tool with one public founder gets one row.

Email reality: verified public emails are rare (often <2% for small/indie tools). True deliverability verification needs SMTP tooling + clean IPs. So mark inferred patterns honestly and tell the user to validate before outreach — LinkedIn + a named founder is usually the reliable channel.

Step 7 — Synthesize

• A normalized master CSV per grain (single import).
• A short README/index with headline counts, per-dataset breakdown, method, and the caveats above.
• Optional: a self-contained, filterable HTML dashboard (search + category + icp_fit filters + distribution bars) — see the dataviz/artifact-design guidance if available.
• A Tier-1 shortlist: icp_fit = High + has a named decision-maker + LinkedIn, sorted by fit strength → the ready-to-work outreach list.

scripts/rollup.py normalizes many per-source CSVs (different schemas) into the two master grains and prints integrity stats.

Pitfalls

Combining with other skills

• parallel-agent-fanout — run every discovery + contact-pull wave concurrently; this skill is the what, that skill is the how at scale.
• deep-research — deeper multi-source verification of a specific high-value target.
• company-legal-reputation-research — vet a shortlisted prospect before outreach.
• domain-email-enumeration — upgrade inferred-pattern emails toward verified for the Tier-1 shortlist.
• competitor-identification — seed Tier-1 source-code search with the right competitor embed strings.
• intelligence-dossier — promote a top prospect into a full standing dossier.
• linkedin-activity-intelligence / era-validated-linkedin-analysis — enrich the decision-maker contacts.