Claude Code becomes a GTM second brain when you put your go-to-market knowledge into the files it loads on its own: a short CLAUDE.md for standing rules, imported context files for ICP and voice, skills for repeatable workflows, subagents for isolated research, auto memory for lessons learned, and an MCP connection for live data. Each piece loads at a different time, which keeps context small.

This post is the setup, folder by folder, using Claude Code's actual mechanisms as documented in September 2026. For the concept of a GTM brain, see what a GTM brain knows and how it decides. For a full CLAUDE.md, see the CLAUDE.md template for GTM teams.

What does a GTM second brain in Claude Code consist of?

The trick is matching each kind of knowledge to the Claude Code feature that loads it at the right moment. Everything that loads every session costs context in every session.

What does the project folder look like?

gtm-brain/
├── CLAUDE.md                  # standing rules, imports, project map
├── .mcp.json                  # Databar and other MCP servers, shared with the team
├── context/
│   ├── icp.md                 # segments, titles, exclusions
│   ├── voice.md               # how we write
│   ├── positioning.md         # what we say vs competitors
│   └── accounts/              # one file per key account, read on demand
├── .claude/
│   ├── rules/
│   │   └── campaign-copy.md   # loads only when editing campaigns/**
│   ├── skills/
│   │   ├── account-brief/SKILL.md
│   │   └── weekly-pipeline-review/SKILL.md
│   └── agents/
│       └── account-researcher.md
├── campaigns/                 # one folder per campaign: list, copy, results
└── data/                      # CSV inputs and outputs

gtm-brain/
├── CLAUDE.md                  # standing rules, imports, project map
├── .mcp.json                  # Databar and other MCP servers, shared with the team
├── context/
│   ├── icp.md                 # segments, titles, exclusions
│   ├── voice.md               # how we write
│   ├── positioning.md         # what we say vs competitors
│   └── accounts/              # one file per key account, read on demand
├── .claude/
│   ├── rules/
│   │   └── campaign-copy.md   # loads only when editing campaigns/**
│   ├── skills/
│   │   ├── account-brief/SKILL.md
│   │   └── weekly-pipeline-review/SKILL.md
│   └── agents/
│       └── account-researcher.md
├── campaigns/                 # one folder per campaign: list, copy, results
└── data/                      # CSV inputs and outputs

gtm-brain/
├── CLAUDE.md                  # standing rules, imports, project map
├── .mcp.json                  # Databar and other MCP servers, shared with the team
├── context/
│   ├── icp.md                 # segments, titles, exclusions
│   ├── voice.md               # how we write
│   ├── positioning.md         # what we say vs competitors
│   └── accounts/              # one file per key account, read on demand
├── .claude/
│   ├── rules/
│   │   └── campaign-copy.md   # loads only when editing campaigns/**
│   ├── skills/
│   │   ├── account-brief/SKILL.md
│   │   └── weekly-pipeline-review/SKILL.md
│   └── agents/
│       └── account-researcher.md
├── campaigns/                 # one folder per campaign: list, copy, results
└── data/                      # CSV inputs and outputs

gtm-brain/
├── CLAUDE.md                  # standing rules, imports, project map
├── .mcp.json                  # Databar and other MCP servers, shared with the team
├── context/
│   ├── icp.md                 # segments, titles, exclusions
│   ├── voice.md               # how we write
│   ├── positioning.md         # what we say vs competitors
│   └── accounts/              # one file per key account, read on demand
├── .claude/
│   ├── rules/
│   │   └── campaign-copy.md   # loads only when editing campaigns/**
│   ├── skills/
│   │   ├── account-brief/SKILL.md
│   │   └── weekly-pipeline-review/SKILL.md
│   └── agents/
│       └── account-researcher.md
├── campaigns/                 # one folder per campaign: list, copy, results
└── data/                      # CSV inputs and outputs

Commit everything except secrets and personal notes. A CLAUDE.local.md at the root holds personal preferences and should be in .gitignore. Auto memory lives outside the repo, in your home directory.

What goes in CLAUDE.md, and what stays out?

Anthropic's guidance is to keep each CLAUDE.md under about 200 lines, because longer files cost context and reduce adherence. Put in facts Claude needs in every session. Move procedures to skills and area-specific rules to .claude/rules/.

# GTM brain: Acme

## What this project is
Outbound and account research for Acme (B2B billing software for SaaS).

## Always read
@context/icp.md
@context/voice.md

## Project map
- campaigns/: one folder per campaign
- context/accounts/: account notes, read the relevant file before any account work
- data/: CSV inputs and outputs, never edit inputs in place

## Rules
- Use Databar (MCP) for company, contact and email data.
- Show a credit estimate and ask before enriching more than 50 rows.
- Never write to the CRM or send email. Produce files for review.
- When I correct an approach twice, propose a CLAUDE.md change

# GTM brain: Acme

## What this project is
Outbound and account research for Acme (B2B billing software for SaaS).

## Always read
@context/icp.md
@context/voice.md

## Project map
- campaigns/: one folder per campaign
- context/accounts/: account notes, read the relevant file before any account work
- data/: CSV inputs and outputs, never edit inputs in place

## Rules
- Use Databar (MCP) for company, contact and email data.
- Show a credit estimate and ask before enriching more than 50 rows.
- Never write to the CRM or send email. Produce files for review.
- When I correct an approach twice, propose a CLAUDE.md change

# GTM brain: Acme

## What this project is
Outbound and account research for Acme (B2B billing software for SaaS).

## Always read
@context/icp.md
@context/voice.md

## Project map
- campaigns/: one folder per campaign
- context/accounts/: account notes, read the relevant file before any account work
- data/: CSV inputs and outputs, never edit inputs in place

## Rules
- Use Databar (MCP) for company, contact and email data.
- Show a credit estimate and ask before enriching more than 50 rows.
- Never write to the CRM or send email. Produce files for review.
- When I correct an approach twice, propose a CLAUDE.md change

# GTM brain: Acme

## What this project is
Outbound and account research for Acme (B2B billing software for SaaS).

## Always read
@context/icp.md
@context/voice.md

## Project map
- campaigns/: one folder per campaign
- context/accounts/: account notes, read the relevant file before any account work
- data/: CSV inputs and outputs, never edit inputs in place

## Rules
- Use Databar (MCP) for company, contact and email data.
- Show a credit estimate and ask before enriching more than 50 rows.
- Never write to the CRM or send email. Produce files for review.
- When I correct an approach twice, propose a CLAUDE.md change

The @context/icp.md lines are imports: Claude Code expands them into context at launch. Use imports for the few files that apply to almost every task. Leave larger or rarer files, like per-account notes, as plain paths Claude reads when needed. Run /context in a session to see which memory files actually loaded, and /init if you want Claude to draft a first CLAUDE.md for you.

How do you keep account notes and copy rules from bloating context?

Two patterns. First, account notes stay out of CLAUDE.md entirely. One markdown file per account in context/accounts/, and a rule that says to read the relevant file before account work. Claude pulls in one file, not fifty.

Second, rules that only matter for one kind of work get a paths scope:

---
paths:
  - "campaigns/**/*.md"
---
# Campaign copy rules
- First line references a verified fact about the account, with its source.
- No more than 90 words. One call to action.
- Never claim results we can't show.

---
paths:
  - "campaigns/**/*.md"
---
# Campaign copy rules
- First line references a verified fact about the account, with its source.
- No more than 90 words. One call to action.
- Never claim results we can't show.

---
paths:
  - "campaigns/**/*.md"
---
# Campaign copy rules
- First line references a verified fact about the account, with its source.
- No more than 90 words. One call to action.
- Never claim results we can't show.

---
paths:
  - "campaigns/**/*.md"
---
# Campaign copy rules
- First line references a verified fact about the account, with its source.
- No more than 90 words. One call to action.
- Never claim results we can't show.

Saved as .claude/rules/campaign-copy.md, this loads only when Claude reads files under campaigns/. Rules without a paths field load every session, like CLAUDE.md.

Which workflows should become skills?

A workflow earns a skill when you have run it at least a few times and the steps have stopped changing. Good GTM candidates: account briefs, list building for a known segment, weekly pipeline reviews, post-call summaries. A skill is a folder with a SKILL.md; the folder name becomes the command.

---
name: account-brief
description: Build a one-page brief on a target account. Use when the user asks to research, brief or prep for a specific company.
---

1. Read context/icp.md and context/accounts/<domain>.md if it exists.
2. Use the Databar MCP "Lookup company data" waterfall on the domain for
   headcount, industry, location and funding.
3. Run the job_postings_getter waterfall and note open Sales, RevOps and
   Marketing roles.
4. Write context/accounts/<domain>

---
name: account-brief
description: Build a one-page brief on a target account. Use when the user asks to research, brief or prep for a specific company.
---

1. Read context/icp.md and context/accounts/<domain>.md if it exists.
2. Use the Databar MCP "Lookup company data" waterfall on the domain for
   headcount, industry, location and funding.
3. Run the job_postings_getter waterfall and note open Sales, RevOps and
   Marketing roles.
4. Write context/accounts/<domain>

---
name: account-brief
description: Build a one-page brief on a target account. Use when the user asks to research, brief or prep for a specific company.
---

1. Read context/icp.md and context/accounts/<domain>.md if it exists.
2. Use the Databar MCP "Lookup company data" waterfall on the domain for
   headcount, industry, location and funding.
3. Run the job_postings_getter waterfall and note open Sales, RevOps and
   Marketing roles.
4. Write context/accounts/<domain>

---
name: account-brief
description: Build a one-page brief on a target account. Use when the user asks to research, brief or prep for a specific company.
---

1. Read context/icp.md and context/accounts/<domain>.md if it exists.
2. Use the Databar MCP "Lookup company data" waterfall on the domain for
   headcount, industry, location and funding.
3. Run the job_postings_getter waterfall and note open Sales, RevOps and
   Marketing roles.
4. Write context/accounts/<domain>

For skills that spend credits or touch other systems, add disable-model-invocation: true to the frontmatter so they only run when you type the command. The best Claude Code skills for GTM has ten more recipes, and how to build Claude Code skills for GTM covers writing descriptions that trigger reliably.

When should a subagent do the work?

Use a subagent when a task would flood the main conversation with material you don't need afterwards: reading ten competitor pages, scanning a long call transcript, researching several accounts in parallel. The subagent works in its own context window and returns a summary.

---
name: account-researcher
description: Researches one company from public sources and returns a short, sourced summary. Use for account research that needs many web pages.
tools: Read, Grep, Glob, WebFetch, WebSearch
model: sonnet
---
You research one company at a time. Return at most 15 bullet points,
each with a source URL. Say "not found" rather than guessing

---
name: account-researcher
description: Researches one company from public sources and returns a short, sourced summary. Use for account research that needs many web pages.
tools: Read, Grep, Glob, WebFetch, WebSearch
model: sonnet
---
You research one company at a time. Return at most 15 bullet points,
each with a source URL. Say "not found" rather than guessing

---
name: account-researcher
description: Researches one company from public sources and returns a short, sourced summary. Use for account research that needs many web pages.
tools: Read, Grep, Glob, WebFetch, WebSearch
model: sonnet
---
You research one company at a time. Return at most 15 bullet points,
each with a source URL. Say "not found" rather than guessing

---
name: account-researcher
description: Researches one company from public sources and returns a short, sourced summary. Use for account research that needs many web pages.
tools: Read, Grep, Glob, WebFetch, WebSearch
model: sonnet
---
You research one company at a time. Return at most 15 bullet points,
each with a source URL. Say "not found" rather than guessing

Save it to .claude/agents/account-researcher.md. Two things to know: the tools field restricts what the subagent can do, and subagents don't load the main conversation's auto memory, so put anything they must know in their own instructions. Claude Code sub-agents for GTM covers when the overhead is worth it.

How does memory work across sessions?

Claude Code has two memory systems, and the difference matters for a second brain:

When you tell Claude "remember that we never target agencies", it saves that to auto memory. If the whole team should follow it, say "add this to CLAUDE.md" instead. Run /memory to browse both, open files in your editor, or switch auto memory off. Review auto memory now and then: it holds what Claude inferred, which can go stale when your ICP changes.

How does live data get into the second brain?

Notes about accounts go stale; headcount, open roles and contacts change weekly. Connect a data source through MCP so skills can fetch current facts instead of trusting old notes. To share the connection with your team without committing a key, add a project-scoped .mcp.json that reads the key from each person's environment:

{
  "mcpServers": {
    "databar": {
      "type": "http",
      "url": "https://mcp.databar.ai/mcp",
      "headers": {
        "Authorization": "Bearer ${DATABAR_API_KEY}"
      }
    }
  }
}

{
  "mcpServers": {
    "databar": {
      "type": "http",
      "url": "https://mcp.databar.ai/mcp",
      "headers": {
        "Authorization": "Bearer ${DATABAR_API_KEY}"
      }
    }
  }
}

{
  "mcpServers": {
    "databar": {
      "type": "http",
      "url": "https://mcp.databar.ai/mcp",
      "headers": {
        "Authorization": "Bearer ${DATABAR_API_KEY}"
      }
    }
  }
}

{
  "mcpServers": {
    "databar": {
      "type": "http",
      "url": "https://mcp.databar.ai/mcp",
      "headers": {
        "Authorization": "Bearer ${DATABAR_API_KEY}"
      }
    }
  }
}

Each teammate sets DATABAR_API_KEY in their shell, and Claude Code asks them to approve the project server the first time. The Databar MCP gives the agent access to 100+ data sources, waterfalls for emails and company data, and tables for reviewing results. Lookups that return nothing aren't charged, and in a waterfall only the provider that returns data is billed. The Databar MCP overview lists the tools, and Owler is one of the providers behind the company data waterfall used in the account-brief skill above.

How is this different from a Notion wiki or Obsidian vault?

These aren't exclusive. Many teams keep the wiki for company-wide docs and point Claude Code at the few pages that drive GTM work.

Where does a Claude Code second brain break?

• CLAUDE.md drift. Rules that no longer match how you sell. Review it when your ICP or offer changes, and delete as often as you add.

• Contradictions. Two rules that conflict get followed inconsistently. Check CLAUDE.md, rules files and auto memory together.

• Skill sprawl. A skill for every prompt makes descriptions overlap and the wrong one trigger. Merge skills that do half a job each.

• Machine-local memory. Auto memory doesn't follow you to a teammate's laptop or a cloud session. Anything the team relies on belongs in committed files.

• Stale facts. Account notes written months ago. Have skills refresh key facts from live data before using them.

For how this fits a wider team rollout, see Claude Code for GTM teams and context engineering for GTM teams.

FAQ

What is the difference between CLAUDE.md and a skill?

CLAUDE.md holds facts and rules that apply to every session and always loads. A skill holds the steps for one workflow and loads only when you invoke it or Claude decides the task matches its description.

Does Claude Code remember things between sessions on its own?

Yes, through auto memory, which is on by default. Claude saves notes from your corrections and preferences into a memory folder for the project and loads the index at the start of each session. It is local to your machine.

Can a whole GTM team share one second brain?

Yes, for everything in the repository: CLAUDE.md, context files, rules, skills, subagents and .mcp.json. Each person's auto memory and CLAUDE.local.md stay personal.

How long should CLAUDE.md be?

Under about 200 lines. If it grows past that, move procedures into skills and area-specific rules into .claude/rules/ with path scopes.

Build yours this week

Start with a 30-line CLAUDE.md, one context file and one skill you already run by hand, then connect live data so the skill works from current facts. Start free with the 14-day trial and 100 credits, or book a founder demo to set up the data side with us.