Creating a Claude Skill for Your Sales Process: Here's How It Works

Some links on this page are partner links. They cost you nothing extra, and I only recommend tools I use myself.>

Direct Answer: To create a Claude Skill, you make a folder with one SKILL.md file: a YAML header with a name and description (the trigger), followed by a Markdown instruction for the complete process. Claude automatically loads the Skill once the task matches the description. Build time: 30 to 60 minutes for a simple Skill, half a day for a complete sales workflow.

Spoiler: most people who use Claude daily use only a small part of its capabilities.

Not because they are lazy. But because Claude's default interface (you ask a question, you get an answer) makes it almost impossible to see what comes next.

What comes next: Skills.

A Skill is the difference between "I use Claude" and "Claude works for me." Not as a chatbot you direct, but as a system that executes your workflows: consistently, automatically, exactly as you want it, without you ever having to explain it again.

In this article, I'll show you how I captured an entire meeting workflow (preparation, processing, and follow-up) as a single Claude Skill. And how you can do it yourself. For skills per department, read What are Claude skills?.

Quick Take: A Skill is not a magic script. It's an onboarding document for Claude, so detailed that it knows exactly what you mean without you having to say it every time.

Why Isn't a Good Prompt Enough?

If you use Claude regularly for the same type of task, you'll eventually notice something: you keep typing the same instructions. Not literally, but the essence is always the same.

"Read the HubSpot notes for this contact and create a briefing. Use the same four questions as last week. Keep it concise, two paragraphs. Pay attention to the BuiltWith data."

Every time. Every conversation. With minor variations that result in small variations in output, and you're never sure if it looks exactly like the last time.

That's the problem a Skill solves. You write the instruction once, exactly as you want it. Claude loads it automatically when the task demands it. You only provide the context (who, when), and the rest is identical to the hundred times before.

What is a Claude Skill Technically?

A Skill is a folder. Inside it is one required file: SKILL.md.

That file starts with a short YAML header: the Skill's name and a description of up to 200 characters. That description is critical because Claude uses it to decide when the Skill is relevant. Write it as a trigger, not as a description.

Then follows the actual instruction in plain Markdown, as detailed as you want. The more detailed, the more consistent the output.

The clever part of the architecture: when starting a session, Claude only reads the name and description of each Skill (a few lines, almost no context). Only when a task matches the description does Claude load the full instruction. So you can have ten Skills side-by-side without them interfering.

Skills work in Claude.ai (Pro, Max, Team, Enterprise), Claude Code, and via the API. Team members who get access to the same Skill automatically work in the same way.

What Does a Complete Sales Skill Look Like?

This is the Skill I use for every client conversation. I'll explain it in detail here, not for you to copy it, but to show how specific a good Skill needs to be.

Name: Meeting Preparation and Follow-up

Description (for the YAML header):

Use for preparing for a client meeting, or for processing a meeting afterward. Asks for the name or email address of the conversation partner.

That description is the trigger. As soon as I type "prepare the meeting with [name]" or "process this afternoon's meeting with [company]", Claude recognizes that this Skill applies.

Part 1: Preparation

The instruction precisely describes the steps Claude must follow:

1. Find the contact in HubSpot using the HubSpot connector. Use the email address or name as the search query.
2. Retrieve the complete contact history: notes, calls, deals, and email history from the past 12 months.
3. Read the BuiltWith properties on the contact card (property group: "BuiltWith"): what technologies does this company use?
4. Answer exactly these four questions, in this order, in this structure:

Last Time: What was discussed in the most recent contact moment? What was the agreed-upon next step? If there is no note: state that explicitly.

Current Situation: What deal stage is this contact in? Is there recent activity (website, email, call)? What does that suggest about their urgency or buying readiness?

Opportunities: Based on the technology stack and company size: is there a logical expansion or upsell opportunity? Be specific, name the tool or theme.

Opener: Write one concrete opening sentence for the conversation. Not a greeting, but a substantive sentence that connects to something current with this company. Use the BuiltWith data or recent activity as a starting point.

1. Formatting: maximum three paragraphs, no bullets, readable in two minutes.

Part 2: Processing After the Meeting

After the meeting, I use the same Skill, with a different trigger: "process the meeting."

The instruction:

1. Read the most recent Granola note for this contact via the Granola MCP connection.
2. Write an internal note for HubSpot: what was discussed, what are the three most important points, what are the exact action items with owner and deadline. Maximum 200 words. Always use the same structure: Discussed / Action Items / Next Step.
3. Add the note to the correct contact in HubSpot via the HubSpot connector. Ask for confirmation before writing.
4. Write a client-facing follow-up email: two paragraphs, professional yet warm, includes the three discussed points and the agreed-upon next step. End with a concrete call to action for the client. Prepare the email as a draft in Gmail via the Gmail connector. NEVER send automatically.
5. Afterwards, ask: "Would you like me to update the deal stage in HubSpot?"

What Makes a Skill Different from a Good Prompt?

A prompt gives Claude direction. A Skill gives Claude a working method.

The difference lies in three things:

Consistency. Every briefing has the same four questions, in the same order, in the same format. Not because Claude figures it out itself every time, but because the Skill dictates it. If a colleague gets access to the same Skill, they produce identical briefings. That's scalability.

Memory without context. Claude has no memory between sessions. But the Skill does. Everything you've ever learned about what makes a good briefing (the structure, the tone, the boundary between what Claude does independently and what it asks first) is in the Skill. Every new session starts with that knowledge.

The human-in-the-loop boundary. In my Skill, it explicitly states: Claude can write notes to HubSpot independently (after confirmation). Never send emails. That boundary is fixed in the instruction, not in my head, where I might forget it.

Quick Take: How long does it take to build a Skill? For a simple Skill with one workflow: 30 to 60 minutes. For a complex Skill with multiple parts like the one above: half a day, including testing.

How to Build Your Own Claude Skill, Step-by-Step?

Step 1: Describe the process in plain language.

Don't start with the Skill itself. Begin by writing down what you do when you perform this task manually. Which steps, in what order, with what result. Sketch it as concretely as possible, including the small decisions you make automatically.

Step 2: Ask Claude to write the Skill for you.

Paste your process description into Claude and ask: "Write this as a Claude Skill in SKILL.md format. Use YAML frontmatter with a name and description. The description should act as a trigger for when I want to perform this task." Claude will generate the structure. You refine the instruction.

Step 3: Test with a real case.

Use the Skill for a real appointment. See what Claude does and where it deviates from what you expect. Adjust the instruction, not the prompt, but the SKILL.md.

Step 4: Add the boundaries.

Explicitly write in the Skill what Claude may and may not execute independently. What does it ask for confirmation first? What does it always do as a draft? What does it never do?

Step 5: Share the Skill with your team.

Place the Skill folder in a shared environment. Team members who gain access will work in exactly the same way from day one, without training, without a manual.

When Does a Skill Pay Off, and When Not?

Quick Take: A Skill pays off if you perform the same task at least five times a week, the output always needs to have the same structure, or others need to be able to follow the same process.

Build a Skill for:

• Recurring workflows that always follow the same steps
• Processes where consistency is essential: CRM notes, client emails, briefings
• Tasks you want to delegate to colleagues without extensive explanation
• Anything where the boundary between "Claude does it themselves" and "Claude asks first" needs to be deliberate

Do not build a Skill for:

• One-time tasks or research work that is different every time
• Tasks where creativity and variation are desired
• Processes you are still testing: wait until it is stable, then document it

What's My Honest Take?

The first Skill I built was too vague. Too much room for interpretation, too little structure. The output always differed slightly.

The second version was too rigid. So detailed that Claude had no room to notice something unexpected, a signal in the data that the standard briefing wouldn't pick up.

The version I use now is exactly in between: tight enough for consistency, open enough for judgment. You won't find that balance in the first version. Expect two to three iterations before a Skill truly works well.

The Logical Next Step

If you already prepare your calls with AI, as in Prepare for a sales call with AI, this is the logical next step: capture it as a Skill so that it works identically for every conversation, without you or a colleague having to reinitialize it.

Want to learn how to effectively implement Claude for your team from the ground up? Read about a company brain.