Claude Skills 2.0 Explained

Every time AI gets a serious upgrade, a split happens. Some people start getting results that look almost unfair. Everyone else keeps doing the same thing and getting nothing close. Nobody talks about why.

Here's what's actually happening: most people use Claude the same way every time. Type a prompt. Get an output. And every single time, Claude starts from scratch — no memory of how you work, no idea what good looks like for you.

There's a feature built to fix exactly that. It's called Skills. And Anthropic just gave it three massive upgrades in Skills 2.0. Today we're breaking down all three — and walking through two real builds so you can see the difference firsthand.

What Is a Claude Skill? (And Why Most People Are Leaving AI Gains on the Table)

A skill is not code. It's not an app. It's not even a prompt in the usual sense.

A skill is an instructional manual written in plain English that teaches Claude how to do something your way — your workflow, your criteria, your preferences. Once saved, Claude follows that specification every time, without you needing to explain it again.

Think of a carpenter given a detailed blueprint for building a chair: every measurement, every joint, every finish. The carpenter follows it exactly. That's what a skill does for Claude — it turns your specific expertise into a repeatable, consistent process.

To access Skills: go to Claude.ai → Settings → Capabilities → Skills → Go to customize.

What's New in Claude Skills 2.0 — Three Superpowers Explained

Skills 1.0 had two critical problems that made it feel like a black box:

1. No validation — you'd build a skill, use it, get outputs that seemed okay… but had zero proof it was actually doing what you designed it to do.
2. Silent failures — you'd build a PPT skill, say "build a PPT," and Claude would produce a completely generic one. Your skill never got called, and you had no idea why.

Skills 2.0 solves both problems with three new superpowers:

Superpower 1: Evals — Finally Know If Your Claude Skill Is Working

Evals are automated tests. Instead of manually prompting Claude and hoping for the best, Skills 2.0 runs your skill through a range of realistic scenarios and scores the results.

You can see exactly which situations your skill handles well and which ones it falls apart on. Before Skills 2.0, every skill you built was a leap of faith. Evals give you data. That's a fundamental shift — from gut feel to measurable quality.

Superpower 2: Benchmarks — Prove Your Skill Still Adds Value After Model Updates

Here's the question almost nobody was asking: when Claude gets an upgrade, does your skill still matter?

Your skill was built to fill gaps in what the old model could do. But what if the new model already produces great output on its own? Is your skill still adding value — or is it getting in the way?

Think of it this way: you upgrade your kitchen with a new oven and better equipment. Does your old recipe still work on the new setup? Maybe. Maybe not. You need to test.

Benchmarks run the same brief through two parallel tracks — with your skill active and without it — then put the outputs side by side. You get concrete proof of the gap your skill creates. Not a theory. Not a chart. An actual before/after comparison that either confirms your skill is earning its keep or reveals it's been superseded.

Superpower 3: Triggers — Make Sure Claude Actually Invokes Your Skill

This one is simple but high-stakes. When you say "build a PPT," does your PPT skill actually activate? In Skills 1.0: not always. You'd get generic output and have no way to know your skill was never called.

Triggers give you visibility. You can see whether your skill is activating at the right moments — and if it isn't, you adjust the trigger description until it fires exactly when you need it.

The Two Types of Claude Skills: Skill Booster vs Encoded Reference

Before you build anything, there's one more thing to understand — because it changes your entire approach.

Two different problems. Two different solutions. Knowing which type you're building before you start is the difference between a skill that works and a vague instruction that doesn't.

Live Demo 1: Building a Website Copy Generator Skill (Skill Booster)

When Claude writes website copy from a basic brief, it produces generic, templated output. The kind that converts no one. This is a genuine capability gap — not a missing preference, but a missing process.

The fix: build a Website Copy Generator skill that encodes specific copy principles, a page structure, a design system, and conversion rules refined from years of real marketing work. Not "write good copy" — the actual rules that make copy convert.

Before You Build: Use AI to Define the Skill

If you're thinking "I don't have a documented process like that" — here's the trick. Before building any skill, ask Claude:

What are the key decisions and rules I should define for a skill
that generates website copy for landing pages?
List the structure and categories I need to specify.

Claude will pull the structure out of your head for you. Use it as a scaffold, then fill in your actual rules. Once the structure is clear, you build.

The Skill in Action

Here's an example of what a Website Copy Generator skill specification looks like. This is what you'd write in plain English inside the Skills editor:

SKILL: Website Copy Generator

PURPOSE: Generate complete, production-ready landing page copy following
my copy framework and design system.

INPUTS REQUIRED:
- Product/service name
- Target audience
- Primary benefit (single, specific)
- Social proof (if available)
- CTA goal (signup, purchase, waitlist)

COPY PRINCIPLES:
1. Lead with the outcome, not the feature
2. One idea per sentence — no compound claims
3. Social proof goes directly after the hero claim
4. Every section answers: "So what?" for the reader
5. CTA copy = action + outcome (not just "Sign Up")

PAGE STRUCTURE:
Hero → Problem agitation → Solution reveal →
Features as benefits → Social proof block →
FAQ (3 objections) → Final CTA

TONE: Confident, direct, no corporate filler words.
Avoid: "leverage", "utilize", "solutions", "cutting-edge"

Skills 2.0 doesn't just save this — it immediately tests it. It creates a fictional product, runs the skill on it, generates a full landing page, and shows you the output before you commit. You preview it, adjust, and only then save the skill.

Then you run a Benchmark: same brief, same model, same moment — one output with your skill, one without. That side-by-side is when it clicks. The gap between Claude's default output and output that follows your framework is exactly the gap your skill closes.

Live Demo 2: Building an AI News Curator Skill (Encoded Reference / Personal Playbook)

This one is different. Claude can already find AI news. It's decent at it. But a generic news list isn't useful — what's useful is news filtered through your specific curation judgment.

The AI News Curator skill encodes a scoring system and editorial criteria:

SKILL: AI News Curator

PURPOSE: Search for AI news from the last 15 days, score each item
against my criteria, and produce a curated bulletin.

SCORING CRITERIA (rate each item 1–10 per criterion):
1. Broadly applicable — relevant to non-developers, not niche
2. Genuinely interesting — a real AI development, not a product announcement
3. Practical benefit — something the audience can act on in their work

SELECTION: Return top 10 scored news items, reasoning for each score.

FORMAT:
- Ranked news list with scores + brief rationale
- Australia AI Corner: 1 Australia-specific AI story
- Viral AI Tools: 3 emerging tools worth trying
- Curator's Take: 2–3 sentence editorial POV

AUDIENCE: Business users and AI productivity enthusiasts.
Not developers. Filter for business applicability.

Claude isn't being uplifted here — it's being directed. The scoring rubric, the Australia angle, the audience filter — these are the editor's judgment encoded into the skill. Every time you prep your weekly AI news show, you invoke this skill. It handles the curation. You handle the commentary.

That's what an Encoded Reference skill does: Claude had the raw capability. You gave it your judgment.

How to Build Your First Claude Skill — Step-by-Step

1Open Skills in Claude settings

Go to Claude.ai → Settings → Capabilities → Skills → Go to customize. You'll see your existing skills list (empty if you haven't built any).

2Decide: Skill Booster or Encoded Reference?

Ask yourself: is Claude genuinely weak at this task (Skill Booster), or does Claude do it fine but not the way you need it (Encoded Reference)? The answer shapes everything about how you write the skill.

3Use Claude to scaffold your skill

Before writing anything, ask Claude: "What are the key decisions and rules I should define for a skill that does X?" This surfaces the structure you need to fill in — your actual rules, criteria, and preferences.

4Write the skill in plain English — be specific

Paste in your actual rules. Your real process. The criteria you've refined through experience. That's what makes a skill powerful — not the description, but the encoded knowledge inside it.

5Preview the output before saving

Skills 2.0 auto-generates a test output before you commit. Review it. If it's off, adjust the skill specification and regenerate. Only save when you're satisfied.

6Run Evals to find the edge cases

Use the Evals feature to run your skill against a range of realistic scenarios. Note where it succeeds and where it breaks down. Iterate on the specification until the weak spots are covered.

7Run a Benchmark to confirm value

Run the same brief with and without your skill. If the gap between outputs is clear, your skill is working. If there's barely a difference, your skill needs more specificity — or Claude's model may have already caught up.

8Check your Trigger

Verify your skill fires when it should. Test by using the exact phrasing you'd normally use in conversation. If the skill doesn't activate, update the trigger description to match how you actually talk to Claude.

Frequently Asked Questions About Claude Skills 2.0

What is a Claude Skill?

A Claude Skill is an instructional manual written in plain English that teaches Claude how to do something your way — your workflow, your criteria, your style. Unlike a one-off prompt, a skill is saved and reused. Claude follows it automatically every time the trigger fires.

What's the difference between Claude Skills 1.0 and Skills 2.0?

Skills 1.0 had two major blind spots: you couldn't verify whether your skill was working, and you couldn't tell if Claude was even invoking it. Skills 2.0 solves both with three new features: Evals (automated testing), Benchmarks (side-by-side output comparison), and Triggers (visibility into when your skill activates).

What are the two types of Claude Skills?

A Skill Booster fills a genuine capability gap — Claude is weak at something, and your skill uplifts it. An Encoded Reference skill encodes your specific judgment and preferences — Claude can do the task, but not your way. Understanding which type you're building is the most important decision before you start.

What are Triggers in Claude Skills 2.0?

Triggers determine when Claude automatically invokes a skill based on what you say. In Skills 1.0, your skill might never activate — you'd say "build a PPT" and get a generic one with your custom PPT skill completely ignored. In Skills 2.0, Triggers let you see whether your skill is firing at the right moments and tune the trigger wording until it does.

What does Benchmarking do in Claude Skills 2.0?

Benchmarking runs the same brief twice — once with your skill active, once without — and shows you both outputs side by side. This gives you concrete proof that your skill is adding value, not just a gut feeling. It also lets you re-validate your skill every time Claude gets a model update.