Claude Code Skills for GTM Engineers
The reusable building block that turns Claude Code into your team's command center.
What Skills Are and Why They Matter
A Claude Code skill is a markdown file in a directory that tells the agent how to handle a specific task. The directory lives at ~/.claude/skills/<skill-name>/. Inside is a SKILL.md file with a YAML frontmatter block that describes the skill and a body that contains the instructions. When the user's prompt matches the trigger description, Claude Code loads the skill into context and follows the instructions.
The wedge for GTM teams. Without skills, you're typing the same long prompt every time you want to research an account, draft an outreach email, or audit a sequence. With skills, you type "research Acme Corp" and the agent loads the right pattern, calls the right MCP servers, and produces output in your team's standard format. The skill is the institutional memory that survives a context window reset and travels across team members.
The honest framing. Skills are well-structured prompts in a discoverable location, nothing more sophisticated than that. A bad skill that hallucinates the wrong steps will hallucinate them faster. The discipline that matters is writing clear triggers, testing each skill on real input, and version-controlling them as code.
The Skill File Format
Every skill is a directory with at least a SKILL.md file. The frontmatter is the contract between you and the agent:
---name: research-accountdescription: Use when Rome says "research [company]", "look up [company]", or "tell me about [company]". Pulls firmographic data, recent news, and CRM context into a one-page brief.---
The description is what Claude Code reads to decide whether to invoke the skill. Be specific about triggers. Vague triggers cause the model to either over-invoke (loading the skill when it shouldn't) or under-invoke (missing the case it should). Include example phrasings from how your team actually talks.
The body of the file is the instructions. Treat it like a runbook. Step by step. What tools to call, in what order, what format to output. If the skill writes to your CRM or sends an email, spell out the guardrails (approval gates, allowed fields, batch limits) explicitly. A skill that doesn't include guardrails inherits whatever the global guardrails happen to be, which is risky for sensitive operations.
Six Skills Every GTM Team Should Build
Build these in order. Each one earns its place by saving time on a workflow you do at least weekly.
1. research-account. The starter skill. Pull firmographic data, recent funding, hiring signals, tech stack, and CRM history for a company. Output a one-page brief. See the Claude Code sales agent guide for the full research pattern.
2. draft-followup. Generate a follow-up email after a meeting, pulling context from the call notes and the open opportunity record. Output a draft email with a subject line and three subject-line variations. Stage it in the CRM or send to Slack for approval.
3. build-sequence. Take a persona, an ICP segment, and a value prop. Generate a 5-touch outbound sequence with timing, subject lines, and bodies. Output structured for direct paste into Smartlead, Instantly, or Lemlist.
4. audit-sequence. Read a sequence in your sequencer, score each email against deliverability rules (length, link count, spam-trigger words), and surface the weak touches with rewrites. See the cold email deliverability guide for the rules.
5. log-call. Take a meeting transcript or notes, extract the key facts (pain points, next steps, decision criteria, timeline), update the relevant HubSpot or Salesforce fields, and post a summary to a deal channel in Slack.
6. weekly-pipeline-review. Run every Monday morning. Read open opportunities, surface stalled deals (no activity in 14+ days), generate a stack-ranked list with recommended next steps. Output to a Notion page or a Slack DM.
Anatomy of a Working Skill
Here's the structure of a real account-research skill. Frontmatter sets the triggers and the description Claude Code uses for dispatch. The body has five sections.
Context. One paragraph on what the skill does and why. Helps the agent (and you) keep the scope tight when it expands later.
Inputs. What the skill needs to run. Company name or domain. Any optional flags like depth (light vs deep research).
Steps. Numbered and explicit. "Call the HubSpot MCP to get the company record. If it exists, capture deal_count, last_activity_date, and assigned_owner. Call the Clay MCP for firmographics. Call the news search for recent press in the last 90 days."
Output format. Exactly what the result should look like. Field names, ordering, tone, length. If the team has a standard "research brief" template in Notion, the skill outputs in that exact shape.
Guardrails. What the skill does NOT do. "Do not write to the CRM." "Do not send emails." "Do not call paid enrichment vendors if the company is already enriched within the last 14 days." Explicit denial is more reliable than implicit assumption.
Sharing Skills Across the Team
The pattern that works: one repo, one PR per skill change. The repo lives in your team's GitHub org and contains all the team-shared skills. Each engineer clones it into ~/.claude/skills/team-shared/ and symlinks individual skills into their main skills directory.
This gets you four things. Version history on every skill (who changed what, when). Code review on changes (a new skill or a tweak goes through a PR like any other code). Conflict resolution if two engineers change the same skill (the merge conflict surfaces it). Onboarding (a new GTM Engineer clones the repo and has the team's standard tools on day one).
Some teams use a private package manager (npm, PyPI) to publish skills as installable bundles. That's heavier but scales better past 15 to 20 skills. Most GTM teams don't need it until they hit that range.
What Breaks in Practice
Three failure modes show up.
Trigger collisions. Two skills have overlapping triggers and the agent picks the wrong one. The fix: be specific in the description. "research-account" triggers on company research. "research-contact" triggers on person research. Don't use vague triggers like "research X".
Stale instructions. A skill was written when the CRM had different field names. The CRM changed. The skill silently fails or writes to the wrong field. The fix: review skills quarterly. Run them on real data. If output drifts, update or retire the skill.
Over-invocation. The skill gets loaded for tasks it wasn't designed for because the description was too broad. The agent then produces output that doesn't match the task. The fix: tighten the description to include the specific phrasings your team actually uses, and add a "When NOT to use this skill" section if needed.
For the broader Claude Code workflow patterns, see the sales agent guide and the AI SDR build. For the CRM wiring that most skills need, see Claude Code MCP HubSpot and Claude Code MCP Salesforce.
Authoritative References
For the skills format and the agentic skills system in detail, see Anthropic's Claude Code skills documentation. For MCP server setup that skills depend on, see the Claude Code MCP documentation.
Frequently Asked Questions
What is a Claude Code skill?
A skill is a reusable bundle of instructions Claude Code can load on demand for a specific task. It lives as a directory under ~/.claude/skills/ with a SKILL.md file that describes when to invoke it and what to do. When the trigger conditions in the description match the user's request, Claude Code loads the skill into context and follows its instructions. Skills are the primitive that turns Claude Code from a general-purpose coding agent into a GTM-specific tool that knows your team's workflows.
How is a Claude Code skill different from a system prompt or a CLAUDE.md file?
A CLAUDE.md is loaded once at the start of a session and stays in context the whole time. It's where you put persistent project context, voice rules, and global instructions. A skill is loaded conditionally, only when the task triggers it. The benefit of skills: you can have 20 task-specific skills (research an account, write a follow-up email, draft a SOQL query) without all 20 sitting in context for every prompt. The model picks the right skill per task and ignores the rest.
How do I share Claude Code skills across my GTM team?
Put the skills in your team's GitHub repo and have everyone clone them into their ~/.claude/skills/ directory. Or use the team-shared skill location if your Claude Code Team plan supports it (the Team Premium tier does as of 2026). The skills are markdown files, so they version-control cleanly and review through standard PRs. Most teams keep one repo with all shared GTM skills (research-account, draft-followup, build-sequence, etc.) and pin updates monthly.
What's a good first Claude Code skill to build for a GTM team?
An account research skill. It triggers on phrases like 'research [company]' or 'tell me about [company]'. The skill instructs Claude Code to call your CRM MCP server for the company record, the enrichment MCP for firmographics and signals, the web for recent news, and write a one-page brief in a standard format. It's high-value (saves 20 to 40 minutes per use), low-risk (read-only on customer data), and easy to test. Most GTM teams start here and add the higher-stakes skills (write outreach, log calls) once the team trusts the pattern.
Can a Claude Code skill call MCP servers?
Yes. Skills work with the full Claude Code tool surface, including MCP servers, hooks, and subagents. A skill is just instructions for what the agent should do when triggered. If the instructions say 'call the HubSpot MCP server, get the company record, then call the Apollo MCP server', the agent does exactly that. The MCP servers need to be wired separately (in .mcp.json or via claude mcp add). The skill assumes they're available and uses them.
Source: State of GTM Engineering Report 2026 (n=228). Salary data combines survey responses from 228 GTM Engineers across 32 countries with analysis of 3,342 job postings.