Audience Icp Filter

Applies to a list that already exists and needs sorting before anyone contacts it. Produces four labelled buckets, a reason per lead, and reconciled counts.

Every imported list is mostly noise

Colleagues are in it. Competitors are watching. A third of the job titles are unreadable, half the companies are stale, and somewhere in there are the twelve people actually worth a message. The default response — skim it, sort by gut, start sequencing — leaks in the direction that hurts most: someone's own coworker gets a cold pitch.

This skill starts from an existing list. Importing or scraping is a different job with different prerequisites; folding it in makes both slower.

Check the data before asking about the ICP

The instinct is to ask what the ICP is first. Do the opposite: measure what the list actually contains, then ask only about criteria the data can support.

There is no point offering geography filtering on a list where the location field is empty. It doesn't filter anything — it routes everyone into review and calls that a result. The same applies to industry, and to exclusion when there's no company and no email to match on.

Run the coverage check, read which criteria are blocked, and say so before the conversation about the ICP starts. If enrichment is needed, quote what it will cost and get an explicit yes before spending anything. If the team declines, proceed — but name the criteria you dropped and say that exclusion is now best-effort. Filtering on a criterion the data can't support and presenting the result as clean is the worst available outcome, because it looks like work. The field-by-field thresholds and the ICP question set are in references/coverage-and-icp.md.

Two passes, and neither is optional

Pass 1 is deterministic. scripts/build.py pattern-matches seniority and function across the title and, when the title is silent, the bio; applies exclusions uniformly across every identity field; and refuses to emit a result whose bucket counts don't reconcile against the input. Nobody gets lost, and the same list classifies the same way twice. Never hand-sort a list — it's unauditable and it's exactly how colleagues leak through.

Pass 2 is semantic, and it is the agent's job. Patterns can't read meaning. Chief Happiness Officer and Chief Medical Officer both resolve as C-level and neither is a buyer. A competitor nobody thought to list sits quietly in the match bucket. Founder @ Stealth Mode lands in review when it's obviously in ICP.

The point of the skill is that nobody hand-sorts a list. Handing back a fifty-lead review bucket is not a result — it's the original problem with extra steps. Work the queue down; leave only the genuinely ambiguous handful for a human.

Review the queue, not the list

The expensive mistake is re-reading everything in pass 2. On a 250-lead list, lead-by-lead review took eight minutes and changed almost nothing, because most classifications were never in doubt.

Pass 1 returns a bounded queue instead: the ambiguous, the matches inferred from a bio rather than a title, the matches whose company reads like an agency or a freelancer, the leads dropped on a soft geography or industry miss, and the leads dropped on a seniority or function read out of free text rather than off the title. Each carries a flag saying why it's there. On a clean list that's fifteen to twenty-five leads, and it's the whole of pass 2's work. Anything outside the queue was confident enough to trust.

A large review bucket is a signal, not a workload. It usually means the list was never enriched, or the ICP is under-specified — most often that nobody answered whether founders qualify regardless of stated function. Diagnose the cause and say it rather than moving forty leads by hand. references/pass-two-review.md covers the flags and the false-positive patterns.