Most people using Claude every day still copy-paste the same instructions at the start of every session. The same brand voice paragraph. The same output format. The same checklist of things to remember. There is a Claude feature that ends this entirely. It launched on Pro, Max, Team and Enterprise plans, and most users have not built a single one yet. It is called Skills, and the first one you build will change how you use AI for everything else.
A Skill is a folder containing a single Markdown file called SKILL.md. That file teaches Claude how you work, what your standards are, and what steps to follow for a recurring task. You write it once. You enable it. From then on, every conversation you have already knows how you do that task. No copy-paste. No reminders. No drift in quality between Tuesday and Friday.
This guide walks you through building your first Skill in under 30 minutes, with a copy-paste-ready template you can adapt to almost any recurring workflow. No coding required. The whole thing is just text in a file.
What is a Claude Skill?
A Claude Skill is a Markdown file with YAML frontmatter (a name and description) and instructions in plain English that Claude loads automatically when your request matches what the Skill does. Skills replace repetitive prompt copy-paste with a stable, version-controlled workflow that activates in every conversation. Skills are available on Pro, Max, Team and Enterprise plans, and require Code Execution to be enabled in Settings > Feature Previews.
The mental model that makes Skills click: think of a Skill as a colleague's onboarding doc, but for Claude. When a new junior writer joins your team, you do not retrain them every Monday morning. You hand them the brand guide, the workflow doc, and a few example outputs. They read it once and apply it forever. A Skill is the same thing for Claude.
The difference between Skills and the older Custom Instructions or Projects features is reach. Custom Instructions sit globally on your account. Projects scope context to one folder. A Skill is portable, shareable, and triggers automatically based on your request, not based on which window you are in. You can have ten Skills active and Claude picks the right one for the task in front of it.
Anthropic shipped Skills 2.0 in March 2026, adding three things that matter for builders: built-in evals (Claude runs your Skill twice, once with it loaded and once without, and scores the difference), A/B testing (a separate Claude judges which output is better without knowing which used the Skill), and trigger optimisation (an automated process that rewrites your Skill description until it activates reliably). You do not need to use any of these on day one, but they exist when your first Skill needs tuning.
How does Claude know when to use a Skill?
Claude decides whether to load a Skill based entirely on the description field in the YAML frontmatter at the top of your SKILL.md file. The description is matched against the user's request in real time. If the description sounds like the task being asked, the Skill activates. If it sounds vague or generic, the Skill never triggers, no matter how good the instructions inside are.
This single insight is why most first-time Skill builders fail. They write a great instruction set inside the Skill, then write a description like "helps with marketing tasks" and wonder why Claude never uses it. The description is not for humans browsing a list. It is for Claude scanning incoming requests.
The fix is to write descriptions the way users actually phrase requests. If your Skill writes LinkedIn posts, your description should include phrases like "write a LinkedIn post," "draft a LinkedIn update," "create a LinkedIn caption" — the natural variations a real person types. If your Skill builds weekly client reports, include "weekly status report," "client update email," "Friday recap." Cover three to five common phrasings.
A practical rule from Anthropic's own skill-creator meta-skill: the description should sound like the way you would naturally ask for that task in a chat message. Read it out loud. If it sounds like internal documentation, rewrite it. If it sounds like an actual user request, you are close.
How do you create a custom Claude Skill?
To create a custom Claude Skill, you write a single SKILL.md file containing YAML frontmatter (name and description) and the instructions Claude should follow, then upload that file as a Skill in Settings > Capabilities > Skills. The whole structure is plain text. No installation, no environment setup, no command line. The minimum viable Skill is around 30 lines of writing.
Step 1: Define the exact task. Get specific. "Marketing help" is too broad and Claude will never know when to load it. "Write a LinkedIn post that drives profile visits in my voice" is the right shape. The narrower the task, the better the Skill performs.
Step 2: Set the role. Give Claude an identity that matches the expertise. "You are a LinkedIn growth strategist who has helped HK marketers build audiences from 500 to 10,000 followers" is more useful than "You are a writing assistant." The role primes Claude's voice and judgement.
Step 3: Add the context Claude needs. Brand voice rules, audience demographics, jargon to avoid, examples of outputs you love. This is the single biggest determinant of output quality. Skip it and you get generic output. Include three real examples and Claude matches the pattern.
Step 4: Write the instructions as numbered steps. Each step should produce a specific intermediate output that feeds the next step. Step 1 might be "extract the hook from the user's input." Step 2 might be "draft three opening sentences." This forces Claude to think in stages instead of producing a single blob.
Step 5: Specify the output format. Use Markdown headers, bullet points, or a literal template showing exactly what the final deliverable looks like. Show a "good" example and a "bad" example side by side if you want to be ruthless about quality.
What does a working SKILL.md template look like?
A working SKILL.md starts with three-dash YAML frontmatter containing only two required fields, name and description, followed by clear Markdown instructions broken into Role, Context, Steps, and Output Format sections. Below is a complete template you can copy, paste, edit with your own brand details, and upload as your first Skill in under 20 minutes.
Try this prompt template (save it as SKILL.md):
---
name: linkedin-post-writer
description: Write a LinkedIn post that sounds like the user's voice. Triggers when the user says "write a LinkedIn post," "draft a LinkedIn update," "post on LinkedIn," "create a LinkedIn caption," or shares a topic and asks for a LinkedIn version. Use for personal LinkedIn posts targeting HK marketing and tech professionals.
---
# LinkedIn Post Writer
## Role
You are a LinkedIn growth strategist who has helped HK marketing and tech professionals grow audiences from 500 to 10,000+ followers. You write in a clear, opinionated, peer-to-peer voice. Never corporate. Never hype.
## Context (My Voice)
- Audience: HK marketers and operators aged 25 to 45
- Voice: confident, slightly nerdy, no jargon-as-status
- Avoid: emojis, "thrilled to announce," "in today's fast-paced world"
- Length sweet spot: 80 to 150 words
- Always end with a single direct question to drive comments
## Steps
1. Extract the core insight from the user's input. Write it in one sentence.
2. Draft three different opening hooks: a contrarian take, a specific number, a confession.
3. Pick the strongest hook. Build a 3-paragraph post: hook, evidence/example, takeaway.
4. End with a question that makes a reader want to comment.
5. Self-check: does it sound like a real person, or a brand? If brand, rewrite paragraph 2.
## Output Format
Return only the final post text. No preamble. No "here is your post." Just the post itself, ready to paste.
That whole template is roughly 25 lines. Save it as SKILL.md, drag the file into a folder named linkedin-post-writer, zip the folder, and upload the zip in Settings > Capabilities > Skills. From your next conversation onward, asking Claude "write a LinkedIn post about X" loads this Skill automatically.
What are the most common Skill-building mistakes?
The three most common mistakes that stop a first Claude Skill from working are vague descriptions that never trigger, instructions buried in long paragraphs Claude cannot scan, and zero example outputs forcing Claude to guess at quality. All three are fixable in five minutes once you know what to look for.
Mistake 1: The description is for humans, not for Claude. "A Skill that helps with social media" never triggers. Claude is matching against actual user phrasing. Rewrite descriptions in user-language: "Triggers when the user asks to write a LinkedIn post, draft a LinkedIn caption, or create a Twitter thread."
Mistake 2: Instructions are written as essays. A wall of paragraphs is hard for Claude to follow consistently. Convert paragraphs into numbered steps. Each step has one job. Claude executes them in order and produces predictable output. This is the single highest-leverage change you can make.
Mistake 3: No examples. Claude is excellent at pattern-matching when you show what good looks like. Add a "Good Example" and "Bad Example" section near the bottom of your Skill. Two examples often produce a 30 to 50 percent quality lift in testing.
Bonus mistake: Forgetting to test in a fresh chat. Always test your Skill in a brand-new conversation, not the conversation where you built it. The conversation where you built it has all the context already; a fresh chat is where you find out if your description triggers reliably.
How do you test whether a Skill is actually working?
Test a Skill by opening a fresh Claude conversation, asking the task it should handle in three different phrasings, and checking whether Claude loads the Skill in all three cases. If the Skill triggers in only one of three, your description is too narrow. If it never triggers, your description is too generic. The fix is always the description, not the instructions.
A quick five-minute test routine: open a new chat, paste a real task you would actually ask. Look at Claude's response. If the Skill is active, you usually see the output match your specified format and voice immediately. If the output looks generic, the Skill did not load.
If Skills 2.0 evals are enabled on your account, you can run the formal A/B test directly. Claude runs your task with and without the Skill, scores both outputs against criteria you set, and tells you the gap. This is overkill for your first Skill but valuable once you have three or four and want to know which deliver the most lift.
The Skill is working when, three weeks later, you stop noticing it. The output just appears in your voice, in your format, with your standards baked in. That is the moment Skills change from "feature I tried" to "tool I rely on."
What should you build a Skill for first?
Build your first Claude Skill for a task you do at least twice a week, where the output format is well-defined, and where you currently copy-paste the same instructions to Claude every time. Repetition plus consistent format equals the highest payoff. A LinkedIn post writer, a weekly status report builder, or a client meeting summariser are perfect first targets.
The reason "twice a week" matters is that Skills compound. The time you spend building one is amortised across every future use. A Skill you use 100 times next year saves you about 100 prompt setups. A Skill you use twice never pays back the build cost. Pick a high-frequency task.
The reason "well-defined output" matters is that Skills excel at consistency. They are weakest at open-ended creative tasks where you do not yet know what you want. Save those for direct prompting. Skills are for "I know exactly what good looks like and I want it every single time."
One concrete recommendation: start with whatever you most often paste at the top of a new chat. That paste is your first Skill. The fact that you paste it repeatedly is the data telling you it should be a Skill. Build it once this afternoon, and the next conversation that needs it just works.
Conclusion: Skills are the upgrade most people miss
Skills are the upgrade that closes the gap between casually using Claude and actually leveraging it. The build cost is low: 30 minutes for your first one, 10 minutes once you know the pattern. The return is permanent: every conversation about that task, forever, runs at your highest-quality output by default. The frustration of inconsistent prompts disappears.
The reason most users have not built one yet is purely friction. The setting is buried in Feature Previews. The first build feels uncertain. But once one Skill works, building the next four takes an afternoon, and your relationship with Claude changes from "tool I use" to "co-worker who already knows my standards."
懂AI，更懂你 UD相伴，AI不冷。 At UD, we have spent 28 years helping HK businesses turn complex technology into reliable workflows. Skills are exactly that pattern, applied to AI: write the standard once, run it forever.
Build It Right the First Time
You have the template. The next step is mapping it to the actual workflows in your day-to-day work, designing Skills that trigger reliably, and integrating them into your team's AI stack. The UD team will walk you through every step, from your first SKILL.md to a small library of Skills that handles the work you currently copy-paste five times a day.
