Skill — Qualify companies against the client's ICP (the qualification gate)

Use when: a sourcing step produced a set of companies (LinkedIn/job scrape, people/company search, an uploaded list) and you're about to spend credits enriching or sequencing. Run this gate first so only ICP-fit companies pass downstream. This is the qualify stage of the pipeline: source → qualify(ICP) → enrich → sequence → CRM.

Where this runs (Model A)

Qualification is reasoning done by the customer's own Claude (the agent running this skill), reading the client's local ICP.md. No hyreflow server call, no hosted AI model call, no credits for the judgment itself — credits are only spent on the optional firmographic lookups in Step 1 and on enriching the passes.

Inputs

• companies — from the sourcing step; each ideally carries name, domain/website, and any firmographics the source already returned (industry, size, location, the job context).
• ICP.md — the client's Ideal Customer Profile, built by /icp and stored in the working directory (per-client data, not part of this skill). If it's missing, tell the user to run /icp first; do not silently skip the gate.

Step 1 — Hard filters (data rules, NO LLM, ~free)

Read the firmographic constraints from ICP.md (Industry, Company Type, Stage & Size, Geography, and any explicit exclusions) and apply them to the data you already have from the sourcing step:

• Drop companies clearly outside geo / size band / industry / type / exclusions.
• Use fields already present first (job + people-search payloads usually carry company industry, size, location). Only if a decisive hard-filter field is missing do you fetch it with one cheap company enrich — and flag the spend: Apollo().enrich_organization(domain), AiArk().company_search({...}), Lusha().enrich_company(domain=…), ZoomInfo().enrich_companies({...}), or a LeadMagic company enrich. Batch where the adapter supports it.
• Output: survivors → Step 2; dropped → log with the failing rule.

Step 2 — Fuzzy fit (LLM judgment = the customer's Claude, free)

For survivors, judge qualitative fit against ICP.md's signals (mission, product, buyer, "what good looks like"). Pick the path by scale:

• Inline (small N): you (the agent) read each company's site/description + ICP.md and assign a fit tier (A/B/C) or score (0–100) with a one-line reason.
• At scale / parallel: dispatch the company-qualifier agent (it WebFetches each site and scores against a qualification prompt). First derive that weighted qualification prompt from ICP.md; the icp-analysis framework / icp-analyzer agent define the dimensions.
• Threshold: default keep tier A/B (or score ≥ 70); make it configurable per client in ICP.md.

Output

• qualified — fits (with score + reason) → proceed to enrich → verify → sequence → CRM.
• rejected — non-fits, logged with reasons (so the user can audit what was dropped and why).

Only qualified companies move forward — that is the credit-saver.

Where it sits in the pipeline

source (jobs/people/list) → [qualify-against-icp] → enrich → verify → sequence → CRM
                                  ↑ reads client-local ICP.md

Always run the gate before enrichment/outreach, never after — the whole point is to not pay to enrich or sequence non-fits.

Cost discipline (three nested funnels)

1. Hard filters (free, on data you already have) shrink the set.
2. LLM fuzzy-fit runs only on survivors (free — customer's Claude).
3. Enrichment runs only on the passes (credits). Keeps both LLM and API/credit cost minimal.

Notes

• No new adapter, no hyreflow server call — pure agent reasoning + the existing /icp, icp-analysis, company-qualifier, icp-analyzer assets (all customer-side).
• ICP.md is per-client data kept in the client's workspace — keep it out of the hyreflow.ai/ skill.
• Reuse: every sourcing flow (jobs→managers, list-building, inbound) routes through this same gate.