Claude Code for GTM Engineers: The Five-File Setup

How do GTM engineers use Claude Code?

As a workbench for go-to-market jobs: build a list, enrich it, check it, research the accounts, draft outreach, and repeat the jobs that work. Claude Code, Anthropic's agent for the terminal, does each job from a sentence. What makes it reliable is the setup around it, and that setup is plain files in a repository your team can share.

Claude Code needs a Pro, Max, Team, Enterprise or Console account; Anthropic's setup page says the free plan doesn't include it (Anthropic).

1. Connect your data and sending tools

Add each tool as a remote MCP (Model Context Protocol) server. Anthropic's docs call HTTP servers "the recommended option for connecting to remote MCP servers", and the project scope writes them to .mcp.json so you can "Check .mcp.json into version control so everyone on your team gets the same MCP tools" (Anthropic).

$ claude mcp add --transport http --scope project layrcake https://mcp.layrcake.dev/mcp

{
  "mcpServers": {
    "layrcake": { "type": "http", "url": "https://mcp.layrcake.dev/mcp" }
  }
}

Then run /mcp and authenticate; servers that use OAuth open a browser to sign in. We connect Layrcake for the data and outreach loop (search, emails, verification, research, signals, campaigns, replies). Add your CRM's own server alongside it; Anthropic's docs use HubSpot's as an example. The MCP servers roundup covers the other options by job.

Keep the list short. Every server adds tools for Claude to choose between, and overlapping tools make its choices worse.

2. Put your ICP and rules in CLAUDE.md

CLAUDE.md is read at the start of every session, so it is where your go-to-market context lives: what you sell, who you sell to, who to never contact, how you write. A project's ./CLAUDE.md is shared with the team; ~/.claude/CLAUDE.md holds your personal preferences (Anthropic).

One caution from the same page: Claude treats CLAUDE.md "as context, not enforced configuration". Write your rules there, and enforce the ones that matter with permissions (step 4). Our template is below and in the download button under Share.

# Go-to-market context

Claude reads this file at the start of every session. Keep it short and true. Replace everything in [brackets].

## What we sell
[One or two sentences: the product, who it's for, and the result a customer gets.]
Proof we can cite: [one customer result with a number, and where it's published].

## Who we sell to (ICP)
- Companies: [industry], [country or region], [headcount range], [any must-have, e.g. "has a sales team" or "uses Salesforce"].
- People: [titles], and who to skip ([titles that look close but aren't buyers]).
- Never contact: [existing customers, partners, competitors, or a list file in this repo].

## How we prospect
- Find people with the data tools, never from memory. Never guess an email address.
- Verify every email before it goes on a list. Drop invalid and unknown. Keep catch-all only in small numbers and say how many.
- Research before writing: one dated fact per company, with its source link. No fact, no personalised opener.
- Every request has a credit budget. Check each call's cost before making it and stop before the budget runs out. Default budget: [100] credits.
- Report what each step cost, using the charge each tool returns.

## How we write
- First email under [120] words. One problem, one result, one yes/no question. No meeting link in email 1.
- Follow-ups in the same thread, under [60] words.
- Sign as [name], [role].
- Banned: exclamation marks, "hope this finds you well", "quick question", invented facts.

## What needs a human
- Never launch a campaign, buy sending capacity or send a reply without the user's explicit yes in this session.
- Stop at a draft and show: the list, the openers with sources, the emails, and the total credits.

3. Turn repeat jobs into skills

When you run the same job twice, make it a skill. A skill is a SKILL.md file in .claude/skills/<name>/; Claude sees each skill's description all the time and loads the full instructions only when it uses one, so a long method "costs almost nothing until you need it" (Anthropic). Commit the folder and the team gets the skill too.

Two skills we have published, both free to copy: prospect, which finds people, checks fit, verifies emails and drafts a sequence (6 ready-to-send people for 84 credits in our run), and cold-email, which writes sourced openers and a three-step sequence (7 sendable addresses and 3 sourced openers for 78.75 credits).

For a skill that sends or spends, add disable-model-invocation: true to its frontmatter so it only runs when you call it by name. Anthropic suggests exactly this for workflows "with side effects", naming /send-slack-message as an example.

4. Make anything that sends or spends ask first

Permission rules are the guardrail that actually holds. Claude Code checks them in order, "deny, then ask, then allow", and MCP tools are matched by name, such as mcp__github__get_* (Anthropic). Commit .claude/settings.json and the rules apply to everyone in the repo.

This example lets the research tools run freely and makes every tool that creates, sends or buys ask:

{
  "permissions": {
    "allow": [
      "mcp__layrcake__find_leads",
      "mcp__layrcake__find_companies",
      "mcp__layrcake__enrich",
      "mcp__layrcake__verify_email",
      "mcp__layrcake__research_company",
      "mcp__layrcake__get_operation"
    ],
    "ask": [
      "mcp__layrcake__create_campaign",
      "mcp__layrcake__launch_campaign",
      "mcp__layrcake__setup_sending",
      "mcp__layrcake__reply_to_thread"
    ]
  }
}

Layrcake adds a second gate on the server side: launch_campaign and setup_sending answer first with a preview and an approval link, and nothing sends or is bought until a signed-in person approves. A local rule and a server-side approval together mean no single setting can send email by accident.

5. Run jobs on a schedule

Anything you can ask interactively, you can run headless. Pass -p with the prompt and list the tools it may use (Anthropic):

$ claude -p "Use the prospect skill: 10 heads of sales at UK fintechs with 50 to 200 staff. Budget 120 credits. Stop at a draft and write the result to out/prospects.md." \
  --allowedTools "mcp__layrcake__find_leads,mcp__layrcake__enrich,mcp__layrcake__verify_email,mcp__layrcake__find_posts,Write"

For a weekly run, Anthropic's GitHub Action runs Claude Code in your repository's workflows, "including a cron schedule" (Anthropic). Routines, in research preview, schedule jobs from Anthropic's cloud on paid plans.

Two things change without a person at the keyboard. OAuth sign-in can't run in -p mode, so headless servers need a key passed as a header, kept in your CI secrets. And project servers load without the usual approval prompt. Keep scheduled jobs to research and drafts, and leave launching to a person.

What does this setup cost to run?

Claude Code comes with paid Claude plans, sharing their usage limits. The data costs depend on your tools. With Layrcake, our logged runs cost 31 credits ($0.31) for one verified sales leader at each of five accounts, 84 credits for six fully drafted prospects, and 232.2 credits to read everyone who engaged with a 563-person LinkedIn post, at the prices on the days we ran them. Lookups that find nothing are free.

To connect Layrcake to Claude Code, see the setup page.

Questions people ask