A Claude Skill that takes a seed of ten ideal accounts and a description of why they fit, and uses Clay’s enrichment plus public-signal sources to find a hundred more accounts that look like them. The output is a Clay table ready for outbound or for routing to AEs by territory.

What you’ll need

Clay account, Pro plan or higher (lookalike requires the bigger enrichment surface)
Claude.ai or Claude Code
A seed list of ten to twenty closed-won accounts
A short description of why each seed is a good fit (two sentences each is enough)

Setup

Install the Skill. icp-list-builder.skill exposes extract_seed_signals, propose_filters, and score_candidates.
Drop in the seed list. A CSV with company name, domain, and the two-sentence “why we won” note. The Skill reads these and asks Claude to extract the signal pattern: which industries, which sizes, which tech stacks, which growth indicators.
Generate filter candidates. Claude proposes a set of Clay filters that should narrow the universe to lookalikes. Review and edit; this is the highest-leverage step. A bad filter set produces a thousand wrong companies.
Run the Clay table. Apply the filters. Clay enriches each candidate with the same signals the seed analysis used. Expect five hundred to two thousand candidates depending on how tight the filters are.
Score and rank. The Skill scores each candidate against the seed signature using the rubric implied by the seeds. Top one hundred land in your outbound table; the rest stay parked.

How it works

The interesting work is the seed analysis. Most ICP exercises are too abstract: “mid-market SaaS in fintech.” The seed approach inverts this. Instead of writing a description, you point at ten companies you closed and let Claude reverse-engineer what they have in common. The output is concrete: “all ten have between fifty and three hundred employees, all ten use Stripe, eight of ten have a security and compliance page, six of ten hired a head of revenue in the last twelve months.”

That concreteness becomes the filter set. Clay can filter on most of those signals directly. The score step then quantifies how closely each candidate matches the full pattern.

Watch-outs

Seed selection bias. Ten accounts won by the same AE in the same vertical produces a list that looks like that AE’s territory. Pull seeds across reps, segments, and time.
Public-signal staleness. Hiring pages, tech stacks, and funding signals lag reality by weeks or months. Signals are directional, not authoritative.
Filter over-fitting. A filter set that exactly matches all ten seeds and only the ten seeds is useless. Loosen filters until candidate volume hits five hundred to two thousand.
Clay credit cost. Enrichment-heavy lookalike runs burn credits fast. Run on a sampled candidate set first, then expand.

Stack

Clay — enrichment substrate and filter engine
Claude — seed signal extraction, filter proposal, scoring
Outbound destination — wherever the top hundred go next