Claude GTM Masterclass: Build AI Sales Agents from Scratch (2026)

TL;DR

• This Claude GTM masterclass builds a working AI sales agent in five modules: setup, skills, sourcing, enrichment, and outreach.
• Claude Code is the runtime. The SyncGTM MCP is the data layer. Your CLAUDE.md is the shared brain that keeps every step on-ICP.
• No code required — everything runs from plain-English prompts and real MCP tool calls you can copy from this page.
• The agent sources ICP-matched leads, enriches each with a verified email and firmographics, then drafts signal-based outreach ready to send.
• Full build takes 4–6 hours; a complete sourcing-to-outreach cycle runs in 5–10 minutes after that.

Overview

Most "Claude for GTM" content stops at a demo. It shows you one clever prompt and leaves you to figure out the rest.

This masterclass does the opposite. It walks the full build — from an empty terminal to an AI sales agent that sources, enriches, and writes outreach on its own.

You'll set up Claude Code and the Model Context Protocol (MCP) data layer, install the skills that make Claude fluent in go-to-market, then run each stage of the pipeline against live B2B data through SyncGTM's MCP server.

This is a general-audience build. If you can write instructions in English and copy a command, you can finish it. For a narrower agent-team blueprint once you're done here, read our Claude sales team guide.

What Is a Claude GTM Agent?

A Claude GTM agent is an AI agent, built with Claude Code, that runs go-to-market work end to end — sourcing accounts, finding decision-makers, enriching contacts, and drafting outreach from a single prompt.

The difference from a chatbot is autonomy. A chatbot answers one question at a time. An agent reads a goal, calls tools, writes files, and chains steps together until the job is done.

Three parts make it work. Claude Code is the runtime that reasons and executes. The MCP data layer feeds it live B2B data. A CLAUDE.md context file holds your ICP, tone, and positioning so every output stays consistent.

Without the data layer, the agent guesses. It invents emails and references things that aren't true. That's the single biggest reason GTM agents fail — and the problem this masterclass solves in Module 1. For a deeper look at how this fits the broader shift, see AI in B2B sales.

The Stack: What You Need Before You Start

You need four things. None of them require an engineer, and you can set them all up in under an hour.

The runtime and context are one-time setup. The data and skills are what you'll actually work with day to day. Let's build them in order.

Module 1: Set Up Claude Code + MCP

Start with the runtime. Install Claude Code, then connect the SyncGTM MCP so your agent has a data layer before it does anything else.

Once Claude Code is running, add the SyncGTM MCP server with one command:

claude mcp add syncgtm -- npx -y syncgtm-mcp

Verify the connection: open Claude Code, type /mcp, and confirm SyncGTM appears in the tool list with active status. You should see tools like find_companies, find_people, enrich_person, and verify_email.

Now write your CLAUDE.md. This file lives in your project folder, and Claude reads it automatically on every run. It's the difference between generic output and output that sounds like your team.

Include at minimum:

• Company overview — what you sell, in one paragraph
• ICP — industry, headcount range, target job titles, and hard disqualifiers
• Pain points — the three problems your product solves
• Tone of voice — how your outreach should read
• Differentiators — your top three, stated plainly

Store the CLAUDE.md in a shared repo and treat it as a living document. Every module below reads from it, so time spent here pays off five times over. For a detailed walkthrough of wiring enrichment into this setup, see our Claude Code firmographic data guide.

Module 2: Install Your SyncGTM Skills

A skill is a reusable prompt playbook Claude loads on demand — a saved set of instructions for one repeatable job. Skills turn one-off prompts into a system.

For a GTM agent, you want a skill for each pipeline stage. Each one is a short markdown file that tells Claude exactly how to run that step, which MCP tools to call, and what output to produce.

Keep each skill focused on one job. A skill that tries to source, enrich, and write in one pass produces mush. Chained skills that each do one thing well produce clean, reviewable output.

If you run revenue operations, there's a whole catalog of skills worth stealing — see the 6 best Claude skills for RevOps and the broader best Claude skills for B2B sales.

Module 3: Source Leads That Match Your ICP

Sourcing is where the Claude GTM agent earns its keep — it turns a plain-English ICP description into a scored, ready-to-work list.

Instead of building filters in a UI, you describe the target and let the agent query SyncGTM directly. A minimal sourcing prompt:

Read the ICP from CLAUDE.md.
Use SyncGTM find_companies to source 30 accounts matching:
  Series B SaaS, 50-200 employees, North America,
  hiring for a VP of Sales in the last 90 days.
For each account, use find_people to identify the
  economic buyer and one champion.
Return results to prospects.csv with company, name,
  title, and LinkedIn URL.

The agent calls find_companies to build the account list, then find_people to pull decision-makers inside each one. Everything comes back as structured data Claude parses directly — no scraping, no CSV cleanup.

Start narrow. Run this on 20–30 accounts first and read the output before scaling. If the list drifts off-target, the fix is almost always a tighter disqualifier in your CLAUDE.md, not a new prompt. For more on sourcing depth, see our Claude Code for sales teams guide.

Module 4: Enrich Every Lead With Verified Data

Enrichment is the step that decides whether your outreach lands or bounces. A name and a company aren't enough — you need a verified email and enough context to personalize.

SyncGTM uses waterfall enrichment: it tries multiple data providers in sequence until it finds a result, so hit rates beat any single-provider tool. The agent runs it per record:

Read prospects.csv.
For each prospect:
  1. enrich_person -> full profile + firmographics
  2. find_work_email -> waterfall across providers
  3. verify_email -> confirm deliverable before use
Drop any record where the email fails verification.
Write enriched.csv with email, status, headcount, tech stack.

Two rules make this reliable. First, always run verify_email before an address enters a sequence — an unverified list wrecks your domain reputation. Second, drop the records that fail rather than guessing; a smaller clean list beats a large dirty one every time.

According to McKinsey's State of AI in Sales, AI-enabled sales teams cut non-selling time by 30–40%. Automated enrichment captures a big share of that — it removes the manual data-hunting that eats a rep's morning.

Module 5: Automate Signal-Based Outreach

The final module turns enriched leads into outreach that reads like research, not a mail merge. The trick is signals — the agent references something real and current about each prospect.

Before it writes, the agent pulls fresh signals per contact, then drafts a sequence that opens on the strongest one:

Read enriched.csv.
For each prospect:
  1. Pull signals via SyncGTM MCP:
     check_job_change, linkedin_profile_posts, company_product_launch
  2. Draft Touch 1: open on the strongest signal,
     connect to the ICP pain point in CLAUDE.md
  3. Draft Touch 3: follow-up on a key differentiator
  4. Draft Touch 5: break-up email with a direct ask
Write to outreach.csv: email, touch_number, subject, body.

Teams that lead with a real signal report 5–15% reply rates versus 1–2% for generic templates, per G2's sales engagement benchmarks. The specificity is the whole game — a live job change or funding round gives the agent something honest to say.

One non-negotiable: the agent drafts, a human sends. Review 5–10 emails before importing any batch into Instantly or Smartlead. For prompt templates that book meetings, see our Claude cold email guide.

The Claude GTM Masterclass in Action

Here's the full pipeline running as one motion. Say you sell a developer tool and want to reach newly funded Series B startups.

You give the agent a single outcome prompt:

Source 25 Series B SaaS companies (50-200 staff, North America)
that raised in the last 6 months. Find the VP of Engineering
and one senior IC at each. Enrich and verify emails. Pull a
recent signal per contact. Draft a 3-touch sequence each.
Save everything to /campaign/newly-funded/.

What happens next, without further input:

• Source — find_companies returns 25 funded accounts; find_people pulls two contacts each.
• Enrich — enrich_person and find_work_email fill profiles; verify_email confirms deliverability and drops two bad records.
• Signal — company_product_launch and check_job_change surface a fresh hook per contact.
• Draft — the outreach skill writes 48 personalized emails, each opening on that contact's signal.
• Review — you skim the output, tweak two, and import the rest.

Start to finish, that's under ten minutes of agent time and about fifteen of your review. The same run by hand is a full day. This is exactly the workflow covered in depth in our Claude Code for GTM engineering guide and the wider B2B go-to-market strategy playbook.

Common Mistakes and Best Practices

• Skipping the CLAUDE.md. Without shared ICP context, every module produces generic output. This is the most common reason a Claude GTM agent underperforms — fix the context file first.
• Building one "do everything" agent. A monolithic agent fails on complexity. Chain focused skills — source, then enrich, then write — and pass files between them.
• Sending without verification. Always run verify_email before a sequence. One bounce-heavy batch damages your domain and poisons the next thousand sends.
• Using stale data. A CSV from six months ago produces six-month-old personalization. Live MCP signals are what separate research from spam.
• No human gate. The agent drafts; a person reviews and sends. That review step catches hallucinations before they hit an inbox.

Track reply rates by campaign for the first 30 days. If outreach isn't landing, the signal pull or the personalization prompt is where to tune — not the send volume.

Conclusion

A Claude GTM agent isn't a party trick. It's a system — one runtime, one context file, one live data layer — running your top-of-funnel while your reps focus on conversations that close.

Work the modules in order. Get Claude Code and the MCP connected. Write the CLAUDE.md. Install your skills. Then source, enrich, and write — validating on a small segment before you scale.

The data layer is what makes the whole thing real. SyncGTM MCP connects your agent to verified emails, firmographics, and buying signals, so every module runs on data worth acting on.

Start free on SyncGTM — 50 enrichment credits included, no credit card required. Connect the MCP in five minutes and your Claude GTM agent has a live data layer from module one.