Summary: Firecrawl published a five-step B2B prospecting guide on November 2, 2026, written by Juan Pablo Hernandez, that pairs each step with a paste-ready prompt for ChatGPT, Claude, or any MCP-connected agent. The core argument: AI assistants can plan and draft outreach on their own, but they can’t build a reliable account list without live data. The guide shows how /search, /scrape, and the Alexandria provider library close that gap.
Why a chatbot alone goes stale
The guide’s sharpest point is about data freshness, not prompting technique. Funding rounds, headcount, and job titles change monthly, so a list an LLM builds from training data is outdated the moment you send it. Worse, models will guess email addresses rather than admit they don’t know.
The fix the article proposes is connecting a data layer. The Firecrawl plugin for ChatGPT, the Claude connector, and the Firecrawl MCP server give an assistant three capabilities: /search for current news and pages, /scrape to read a company’s site as clean Markdown, and Firecrawl Alexandria for structured company, people, and funding records from providers like Apollo, People Data Labs, FullEnrich, and Fireround.
I covered what Alexandria changes for GTM agents in an earlier post on people enrichment — this guide is essentially that capability applied end to end.
The five steps, and what data each one needs
The process itself is standard outbound: define your ICP, build the account list, find the buyers, check buying signals, write the first touch. What’s useful is how the guide maps each step to a specific data call:
- ICP definition. Scrape your five best customers’ websites and have the assistant derive shared traits, instead of guessing a profile from intuition.
- Account list. Fireround’s rounds/search finds startups by sector, stage, and raise date; Apollo and People Data Labs cover established companies by industry, size, and region.
- Buyers and emails. Apollo’s people search is free for finding who holds a title; FullEnrich’s work-email lookup runs only on your shortlist, so you pay per resolved contact.
One detail worth trusting: Firecrawl ran a Company Enrichment Benchmark of 200 US companies where Apollo scored 89.3% on fields like headquarters and employee count, and People Data Labs scored 80.6%. That kind of number makes provider selection less of a coin flip.
Agents can run the whole job
The guide goes beyond chat. The same Firecrawl tools work through the MCP server in Grok Bot, Claude Code, Codex, or Cursor, so an agent can build the list, qualify it, and draft openers while you work on something else. Setup is keyless — you sign in through OAuth:
claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp-oauthTwo practical guardrails from the article: ask the assistant to show the Firecrawl credit cost before running a paid search, and Alexandria isn’t part of the keyless tier, so provider calls need a signed-in account. The free plan includes 1,000 credits a month.
One honest limitation
The prompts won’t fix a vague ICP. The guide leans on your existing best customers as the seed for the profile, which means the quality of step one still depends on judgment the AI can’t supply. The takeaway for builders: wire up the data layer first, then treat each prompt as a template you tighten against your own account history rather than a finished pipeline.