B2B Lead Research Agent

Turn "we should find more accounts like these" into a ranked, verified, ready-to-work US prospect list - and outreach copy that survives a spam filter and a compliance review.

Use this skill when someone asks for:

• a target account list for a US B2B product or service
• company research: what a company does, who they sell to, recent triggers
• lead scoring, tiering, deduplication, and list hygiene
• ICP (ideal customer profile) definition from scratch
• cold email and LinkedIn outreach copy for US buyers
• territory/account planning, account selection, "why this account" briefs
• cleaning an existing list before sending

Operating principles (non-negotiable)

1. Company-level research from public business sources. Official company websites, press releases, public filings, public registries, public job postings, public directory listings the company chose to publish.
2. Business contact details published for business use. A company contact page, a press contact, a business role address. Not scraped personal accounts, not leaked databases, not breach dumps, not bought lists.
3. Human in the loop. Research and drafting are automated. Sending is manual - the operator opens the draft, checks it, and sends it.
4. One-to-one by default. Bulk blast to a purchased list is out of scope.
5. Opt-out always wins. One click, forever, no questions asked.
6. Respect the source. No bypassing logins, no scraping behind a paywall or a blocklist, no ignoring robots.txt, no rate-limit evasion.

If a request needs any of these, say no to the request and offer the compliant path instead.

Workflow

1. Define the ICP. templates/icp-worksheet.md - firmographics, triggers, disqualifiers, and the two questions a rep must be able to answer.
2. Discover accounts. Public sources only, see references/research-playbook.md.
3. Research each account. One-page dossier: what they sell, who they sell to, size, funding/stock/leadership signals, a trigger, and one credible reason to contact them now.
4. Normalize and score. scripts/lead_sheet.py dedupes and tiers the list.
5. Draft outreach. templates/outreach-sequence.md - 4-touch, US business email, one clear ask, easy to say yes to.
6. Compliance check. scripts/compliance_check.py on every draft.
7. Hand off for human review. Delivered as a sheet + drafts, never as "send this to 5,000 people".

Quickstart

Score and tier a list:

python3 scripts/lead_sheet.py \
  --accounts accounts.csv \
  --icp icp.json \
  --out scored.csv

accounts.csv needs company,domain,employees,industry,state,country,notes. The script normalizes domains, removes duplicates, applies the ICP weights, tiers accounts A/B/C, and prints a summary.

Lint outreach drafts before sending:

python3 scripts/compliance_check.py --drafts outreach/ --out compliance.json

It checks each draft for the CAN-SPAM elements (real from address, physical postal address, working unsubscribe, non-deceptive subject line), plus required statements, and reports a pass/fail per message.

Output contract

Every research run delivers four artifacts:

Nothing else. No "complete contact database", no unverified email lists.

Research standards for US accounts

• Source every fact. Store the URL and the date researched for each claim.
• One page per account. If it does not fit on a page, it is not research.
• Find the trigger. A hiring spree, a new VP of Sales, a funding round, a new product line, an expansion, a compliance deadline, a competitor change, a tech-stack change. No trigger, no email.
• Write the personalization note. The draft must say which line came from which fact. If you cannot point to it, cut the line.
• Prefer recent, public, specific. The newest press release beats the About page.

Sources that work for US company research

Guardrails (hard rules)

• No scraped personal data. No personal social profiles, no personal email guessing outside a company's published pattern, no data broker dumps, no breach or dump data, no "gmail + firstname.lastname" on private individuals.
• No mass automated sending. This skill never sends. Volume stays low enough that each message can be read by a human first.
• No deceptive subjects. No fake "Re:", no fake "Fwd:", no FW: padding, no misleading sender name, no look-alike domains.
• No bought or scraped lists. Business contacts from a company's own published channels, or from a provider the company is licensed to use.
• No fake personalization. Never claim a fact you did not verify, never invent a mutual connection.
• No scraping against a site's terms or robots.txt. No rate-limit evasion, no credential reuse, no circumvention of paywalls.
• Opt-out suppression is permanent. Keep a suppression list and check it before every send.

Compliance essentials (US)

Full detail in references/us-compliance.md. The short version:

• CAN-SPAM (commercial email): truthful headers, non-deceptive subject, identify the message as an ad where required, physical postal address, working unsubscribe, honor unsubscribe within 10 business days.
• TCPA (SMS and calls): prior express written consent for marketing texts and automated calls. Do not treat a form fill as consent for texts unless it said so.
• State laws: California, Colorado, Virginia, Connecticut, Utah and others require opt-in rather than opt-out for some categories. Opt in everywhere unless counsel says otherwise.
• Platform rules: LinkedIn and major inboxes treat unsolicited automation as abuse and can ban the sending domain. Slow, low-volume, personalized.

Files in this bundle

references/
  us-compliance.md        CAN-SPAM, TCPA, state laws, suppression list rules
  research-playbook.md    ICP definition, discovery, triggers, qualification
  enrichment-sources.md   Public and licensed US company data sources
templates/
  icp-worksheet.md        ICP definition sheet
  company-dossier.md      One-page account brief
  outreach-sequence.md    4-touch US outreach sequence
scripts/
  lead_sheet.py           Normalize, dedupe, ICP-score, tier a list
  compliance_check.py     CAN-SPAM / required-statement linter for drafts

Troubleshooting

Related

• ai-creative-studio — turn each account's trigger into a hook, an ad, and a landing page.
• ai-trading — research public competitors of a target account before you pitch.
• Windows Multi-Tool Suite — one Windows kit that sets up research, content and outbound workflows side by side. The desktop app is a separate download from the publisher's GitHub release, with a published SHA-256 so you can check the file before you unpack it; no skill bundle carries an installer.

Research and drafts only. Nothing in this skill sends anything, and nothing in it bypasses a platform's rules. Sending stays a human decision, and it stays compliant.