Most people treat Claude Skills like fancy system prompts. They dump everything into one blob of instructions, keep the description vague, then wonder why Claude “forgets” to use it. The fix is understanding the architecture and designing for it. (I’ve written about Claude Skills before, yet the question of How To Build them keeps coming up, so here is your tactical guide.)

A Claude Skill is a small, portable package of instructions, along with optional supporting materials, that Claude can load when relevant. The key phrase here is “when relevant.” Skills are designed for progressive disclosure. Claude reaches for resources only if the task needs them. So a Skill isn’t “always-on context.” It’s more like a specialist you can page when the request calls for their expertise. This matters because it shapes how we build them. We’re not creating exhaustive documentation that Claude has to read every time we open a chat. We’re creating targeted workflows that activate precisely when needed. Understanding this distinction prevents most of the frustration people experience with Skills.

Before we build anything, let’s clarify which tool solves which problem. Prompts work for one-off results when reusability doesn’t matter. Skills become essential when we keep repeating a workflow or format across conversations without wanting to paste the same instructions every time. Projects serve as containers for persistent context for a single workstream (a client, a campaign, a product, a class). If you’re typing the same instructions into multiple chats, build a Skill. If you’re gathering research for one specific initiative, create a Project. If you need something once, just prompt it. Projects store a body of work. Skills enforce how work gets done.

Getting Started

Enable the capabilities first. Head to Settings, open Capabilities, and turn on Code execution and file creation (required for many Skill workflows). Then, find Skills and enable the skill-creator example skill. This takes 30 seconds and unlocks everything else. (You’ll need a paid plan for this.)

Now comes the actual work. Before opening a chat with Claude, answer four questions about any Skill we want to create:

• What goes in? What inputs does this workflow need?

• What does Claude do with it? What’s the actual process?

• What comes out? What’s the deliverable?

• What rules must it follow? What makes output good?

These questions force clarity. Vague ideas produce vague Skills. Concrete answers produce Skills that work reliably.

The Easiest Path to Your First Skill

Go back to Settings, Capabilities, Skills, and turn on the brand-guidelines and/or internal-comms example Skills. Click the three dots and use “try in chat.” Tell Claude you want to duplicate this Skill for your own brand. Feed it your style guide or brand guidelines. Or ask Claude to ask you questions for the update: “What materials would help you build this new Skill effectively?”

Claude will ask about your process. This back-and-forth is the actual work. Don’t rush it. The quality of your Skill depends on how clearly you articulate what good looks like. Provide enough detail that someone capable but unfamiliar with your approach could follow it. Be concrete. Instead of “use our brand colors,” specify “always use #2D3142 for primary headings, #BFC0C0 for subheads, #EF8354 for call-to-action elements.”

Watch Claude’s thinking panel (click the dropdown to expand it). You’ll see it reference the skill-creator Skill. This is Claude using its own Skill to build yours, following best practices for structure and formatting. When it’s done, push the button to copy it to your Skills.

Now test like you mean it.

Testing

Don’t do one test and declare victory. Test edge cases. We’re checking three things systematically. Does it trigger without us having to remind it? Check the thinking panel for “Using [Skill name]” appearing automatically. Does output match our specifications? Verify colors, fonts, tone, and structure against our standards. Do edge cases work? See what happens when inputs are missing or requests are unusual.

If it fails, don’t add more text everywhere. Diagnose the type of failure.

When the Skill doesn’t trigger reliably, rewrite the description so it mirrors the words we actually use. Bad description: “Use for internal comms.” Better description: “Use whenever the user asks for a weekly status update, team update, internal announcement, leadership update, internal newsletter section, or any internal email draft. Also use when the user says: status, update, announce, internal, leadership, team note, newsletter.”

If you are having an issue with the Skill, tell Claude, and ask it what would fix the problem.

The Portability Advantage

Most people miss this about Skills: they aren’t locked to Claude’s ecosystem. They’re just files that can be used with any model.

Take your Brand Guidelines Skill and drop it into ChatGPT as a custom GPT or a Gemini Gem. Take any Skill and use it in whatever AI system launches next month. The expertise you spend 60 minutes codifying doesn’t evaporate when Anthropic when a new model becomes the industry standard.

The Setup Tax

First Skill: 45 to 90 minutes. Spend 15 to 20 minutes gathering materials and answering the four questions. Invest 20 to 30 minutes in the conversational build with Claude. Run 10 to 20 minutes of first-round testing. Dedicate 15 to 20 minutes to iteration and refinement.

Second Skill: 30 to 60 minutes. We understand the pattern now.

Third Skill: 20 to 40 minutes. We know exactly what Claude needs to hear.

The quality gate matters here. If we skip gathering good examples, the Skill will be too vague. Claude needs concrete demonstrations of what excellence looks like. If we can’t articulate what makes output good, Claude can’t enforce it. If we don’t test thoroughly, we’ll discover edge cases in production when they’re most annoying. The real value isn’t just time. It’s cognitive load. We stop thinking about formatting, tone rules, banned words, and structure. That mental energy goes to strategy and creativity instead. Skills handle execution patterns. Judgment still belongs to us: what matters, what is interesting, what is worth saying, and what should be cut.

What to Build First

Start with the Skill that removes the most repeated friction. The one where we catch ourselves typing the same instructions for the third time in a week. That’s the Skill worth building. The one that delivers immediate practical value. What we get when this is done well: consistency without repetition, better first drafts that require refinement rather than rescue, and less cognitive overhead spent managing context instead of doing actual work.

That’s the whole game. Codify taste into a reusable playbook, keep triggers explicit, and let the architecture do its job. We can teach Claude our workflows through conversation without writing code or messing with file structures. Skills work across every conversation automatically, which means we stop managing AI context and start managing actual work.

More Resources

I’ve built a growing library of personal Skills, and I’ve bookmarked a few public Skill directories that have saved me more than once. If you tell me what you’re trying to build, I can point you to the most relevant templates and examples. And if you try this guide, I’d love to hear what you made, and what you want to standardize next.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

I’ve also started publishing more frequently on LinkedIn, and you can follow me here

if you’d like to chat further about opportunities or interest in AI, or this newsletter, please feel free to reply.

banner images created with Midjourney.