Lead Enrichment Agent — Testing skill for Claude Code

Autonomous Lead Enrichment Agent

A Python pipeline that takes a list of company domains, crawls their public web presence with a headless browser, cleans the content down to LLM-friendly text, and extracts structured company intelligence (overview, ICP, contacts, leadership, confidence score) using Claude's tool-calling for strict structured output.

How it works

domains  ──▶  crawler.py  ──▶  extractor.py  ──▶  llm_extract.py  ──▶  output.json
            (Playwright:      (BeautifulSoup:      (Anthropic tool-use,
             homepage +        strip scripts/nav,   strict JSON schema
             relevant           collapse to plain    via models.py)
             subpages)          text, cap length)

1. src/crawler.py — Launches headless Chromium via Playwright, fetches the homepage, waits for JS-rendered content to settle, then discovers and fetches up to 5 same-site subpages matching common slugs (about, team, contact, pricing, etc). Handles timeouts, 404s, and bot blocks without crashing.
2. src/extractor.py — Strips , , , , and other boilerplate, preserves mailto: and LinkedIn hrefs inline (so emails/profile links survive text extraction), and caps each page at ~6k characters to bound token spend.
3. src/llm_extract.py — Sends the cleaned text to Claude with a single forced tool call (record_company_intelligence), so the response is always valid structured JSON matching the schema in src/models.py — no regex-parsing of free text.
4. src/main.py — Orchestrates the batch run. Every domain is wrapped in its own try/except at multiple layers, so one failing site (timeout, block, malformed HTML, LLM error) never aborts the rest of the batch — it's recorded with data_confidence_score: 0.0 and an errors list instead.

Setup

git clone 
cd lead-enrichment-agent
python -m venv .venv && source .venv/bin/activ