B2B Lead Hunter

An AI Agent Skill — not a standalone app. It runs inside Claude Code, Cursor, Cline, or Hermes to automate foreign-trade B2B lead research, decision-maker discovery, customs/import signal checks, outreach draft generation, and controlled SMTP sending.

From the team behind b2binsights.io — B2B export intelligence for Chinese manufacturers.

Designed for export sales teams, founders, and B2B sellers who need traceable, verifiable leads — not spam lists. Every company, contact, score, decision maker, and outreach draft is tied to source URLs and structured artifacts. Quality gates override target count.

📺 Watch demo on Bilibili (Chinese walkthrough)

Features

Quick Start

# 1. Clone and install
git clone https://github.com/your-username/b2b-lead-hunter.git
cd b2b-lead-hunter
pip install -r requirements.txt

# 2. Verify setup
python -m py_compile scripts/*.py
python scripts/validate_artifact.py smtp-config templates/smtp-config.example.json

# 3. Run a demo pilot
python scripts/run_hunt.py example/brief.json \
  --out lead-runs/demo/ \
  --queries example/queries.csv \
  --pilot

# 4. See what you got
cat lead-runs/demo/pilot-stats.json

Installation by Platform

This is an AI agent skill — it runs inside an agent, not as a standalone app. Here's how to install it on each platform.

Claude Code

# Option A: Open project directly (CLAUDE.md loads automatically)
cd b2b-lead-hunter
claude

# Option B: Install as a Claude Code project (accessible from anywhere)
claude project add b2b-lead-hunter --path /path/to/b2b-lead-hunter
claude project use b2b-lead-hunter

# Option C: From another directory, reference this repo
claude --project /path/to/b2b-lead-hunter

Once loaded, Claude Code reads CLAUDE.md for context. It uses WebSearch tool for queries, terminal for running Python scripts, and file ops for reading/writing artifacts.

Cursor

# Open the project folder in Cursor
cursor /path/to/b2b-lead-hunter

The rules in .cursor/rules/b2b-lead-hunter.mdc are loaded automatically when Cursor Agent operates on files matching the globs pattern. Use @web() for search queries.

Cline (VS Code)

# Open the project folder in VS Code with Cline installed
code /path/to/b2b-lead-hunter

Cline auto-loads .clinerules at the project root on startup. No additional config needed. Use @web for search queries and the terminal for Python scripts.

Hermes

# Clone the skill into your Hermes skills directory
git clone https://github.com/your-username/b2b-lead-hunter.git hermes-skills/business/b2b-lead-hunter/

Hermes reads SKILL.md as the skill entrypoint. No other setup required. The name field in SKILL.md metadata registers it as b2b-lead-hunter.

All platforms share the same scripts/, references/, and templates/ directories. Platform-specific files (CLAUDE.md, .clinerules, .cursor/rules/, SKILL.md) each contain the same pipeline logic adapted to that agent's capabilities.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                         AI / Agent Layer                         │
│  (search strategy, fit scoring, language choice, approval)       │
└─────────────────────────────────────────────────────────────────┘
                              │
    ┌────────────┬────────────┼────────────┬────────────┐
    ▼            ▼            ▼            ▼            ▼
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌──────────┐ ┌─────────┐
│ Search  │ │ Website │ │ Contact │ │ DM       │ │ Score & │
│ Plan    │ │ Read    │ │ Extract │ │ Discover │ │ Decide  │
│ (AI)    │ │(jina.py)│ │(extract)│ │(enrich)  │ │ (AI)    │
└─────────┘ └─────────┘ └─────────┘ └──────────┘ └─────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                     Deterministic Python Layer                    │
│  (normalize, validate, dedupe, template, export, SMTP)            │
└─────────────────────────────────────────────────────────────────┘

Pipeline Stages

brief.json
  → search-plan.md (AI: research strategy)
  → raw-search.jsonl (normalize search results)
  → candidates.jsonl (dedupe + structure)
  → enriched.jsonl (read websites + extract contacts)
  → leads.jsonl (AI: score + classify)
  → [decision-maker discovery]
  → [customs/import verification]
  → outreach-candidates.jsonl
  → outreach-templates.jsonl
  → outreach-drafts.jsonl
  → outreach-evaluations.jsonl (deterministic quality gate)
  → approved-outreach.jsonl (human approval)
  → sent-log.jsonl (SMTP with hash-locked gates)

Key Design Decisions

• AI + Code separation: AI owns strategy and judgment. Python scripts own deterministic work (reading, extracting, validating, sending). Neither can fully operate without the other.
• JSONL everywhere: Every artifact is JSON Lines with a strict schema enforced by scripts/contracts.py.
• Runtime validation: Every stage validates its output before handoff. No trust between stages.
• Hash-locked sending: Every sent email requires matching hashes from evaluation and approval stages. If the body changes after approval, sending is blocked.
• Compliance-first: Suppression lists, opt-out enforcement, no auto-send to inferred emails, truthful sender identity.

Directory Structure

b2b-lead-hunter/
├── CLAUDE.md              # Claude Code agent instructions
├── .clinerules            # Cline / Codex agent instructions
├── .cursor/rules/         # Cursor agent rules
├── SKILL.md               # Original Hermes skill definition
├── scripts/               # 18 deterministic Python helpers
│   ├── contracts.py       # Runtime contract validators (central dependency)
│   ├── run_hunt.py        # Lightweight orchestrator
│   ├── read_jina.py       # Website reader via Jina API
│   ├── extract_contacts.py
│   ├── normalize_search_results.py
│   ├── dedupe_leads.py
│   ├── export_leads.py
│   ├── enrich_decision_makers.py
│   ├── linkedin_dm_search.py
│   ├── infer_email.py
│   ├── customs_verify.py
│   ├── prepare_outreach.py
│   ├── generate_outreach_templates.py
│   ├── generate_outreach_drafts.py
│   ├── evaluate_outreach_drafts.py
│   ├── send_smtp.py
│   ├── serper_maps.py
│   └── validate_artifact.py
├── references/            # 28 detailed stage guidance files
├── templates/             # JSON schemas and example artifacts
├── example/               # Demo input files
├── _shared/               # Shared SMTP config
├── pyproject.toml
├── requirements.txt
├── Makefile
└── LICENSE

Demo

Run the built-in demo to see real output:

# Install, then:
python scripts/run_hunt.py example/brief.json \
  --out lead-runs/demo/ \
  --queries example/queries.csv \
  --pilot

# View results
cat lead-runs/demo/pilot-stats.json
# → {"query_count": 3, "raw_result_count": ..., "candidate_count": ...}

Check example/ for sample input files you can adapt to your own product and target markets.

Requirements

• Python 3.10+
• requests library (pip install -r requirements.txt)

Optional API Keys

Compliance

This tool is designed for legal, ethical B2B outreach only:

• Collects only publicly available business contact information
• Never bypasses authentication, paywalls, or access controls
• Never generates deceptive identities, fake claims, or unsupported statements
• Respects opt-out requests and regional privacy laws
• Every outreach must include truthful sender identity, source-backed personalization, a low-pressure CTA, and a clear opt-out line
• Auto-sending requires: [user approval → hash-locked evaluation → dry-run → approval hash → suppression check → duplicate check] — all gates must pass

Contributing

Contributions are welcome. Please open an issue or PR.

• Keep the compliance-first design intact
• Preserve the AI/code separation pattern
• Add or update schema validators in contracts.py when adding new fields
• Run make check before submitting

Support

If this project helps your export business, consider supporting it:

• ☕ Buy me a coffee
• 💬 WeChat (tips & feedback)

<p align="left"> <img src="assets/images/WechatIMG1255.jpeg" width="200" alt="WeChat QR" /> <img src="assets/images/WechatIMG1256.jpeg" width="200" alt="Support QR" /> </p>

License

MIT © xiongbojian