Suraj Khaitan

Posted on May 10

🧠 I Tried 100 Claude Skills. These Are The Best.

#ai #claude #cloud #python

From PDF wizards to Slack-GIF generators, I went deep into Anthropic’s new Agent Skills ecosystem — mostly inside Claude Code, where the action really is. Here are the Skills actually worth installing, the Claude Code workflows that have quietly reshaped my dev loop, and the patterns that separate a great Skill from a glorified prompt.

Why I Went Down This Rabbit Hole

When Anthropic dropped Agent Skills in October 2025, my first reaction was: another abstraction layer? My second reaction, after spending a weekend with them inside Claude Code, was: this is how agents actually become useful.

A Skill is deceptively simple — a folder with a SKILL.md file, optional scripts, and reference docs. But the magic is progressive disclosure: Claude only loads what it needs, when it needs it. That means you can hand an agent a 200-page playbook without burning a single token until the moment it’s relevant.

And Claude Code is the place where Skills feel most alive. In the last twelve months it’s gone from a terminal-only experiment to a multi-surface developer environment — terminal, IDE plugin, desktop app, web, iOS, and Slack — powered by Sonnet 4.6 and Opus 4.7, with adoption stories from Ramp, Intercom, Notion, Spotify, Shopify, Figma, Stubhub, and Asana. It’s arguably the fastest-growing AI dev tool on the market right now, and Skills are the layer that turns it from “impressive demo” into “this is how my team ships code.”

So I did the obvious thing. I installed, audited, and stress-tested 100 Skills — from the official anthropics/skills repo, partner Skills, the Agent Skills standard at agentskills.io, and a pile of community contributions on GitHub. Most of the testing happened inside Claude Code, with a few side trips through Claude.ai and the Agent SDK.

This is the shortlist that survived.

TL;DR

Skills ≠ prompts. They’re portable, composable, model-agnostic capability packs.
The best Skills do one thing exceptionally well, lean hard on deterministic code, and have razor-sharp description fields.
My top 10 below cover documents, design, dev workflows, testing, comms, and meta-skills that build other Skills.
Claude Code is the killer host. Plugins, marketplaces, parallel sessions, Routines, and tight Git/Slack integration make it the place Skills shine.
Watch out for trap Skills: bloated SKILL.md files, vague triggers, and Skills that smuggle in untrusted scripts.

A 30-Second Refresher: What Is a Skill?

A Skill is a directory:

my-skill/
├── SKILL.md          # YAML frontmatter + instructions (required)
├── reference.md      # Optional deep-dive context
└── scripts/
    └── do_thing.py   # Optional executable code

The SKILL.md frontmatter only needs two fields:

---
name: my-skill
description: What it does and exactly when Claude should use it
---

At startup, Claude pre-loads only the name and description of every installed Skill. When a task matches, it pulls in the body of SKILL.md. If the body references forms.md, Claude reads that only if needed. Code in the Skill can be executed directly — no token cost for the script body.

This three-tier disclosure (metadata → instructions → bundled assets) is why Skills scale where giant system prompts don’t.

Skills run today across Claude.ai, Claude Code, the Claude Agent SDK, and the Claude Developer Platform, and the format is now an open standard (agentskills.io) for cross-platform portability.

A Detour: Why Claude Code Is Eating the Dev Tool Market

You can’t talk seriously about Skills in 2026 without talking about Claude Code, because that’s where most of the interesting Skill work is happening. A few trends worth naming:

1. It stopped being “just a CLI.”

Claude Code now runs in your terminal, your IDE, a desktop app with parallel task management and visual diffs, the web, iOS, and Slack. The same agent, same context, same Skills — different surface depending on where you happen to be working.

2. Models got dramatically better at long-horizon coding.

Sonnet 4.6 is the everyday workhorse — fast, cheap enough to keep multiple instances running in parallel. Opus 4.7 is the heavy lifter for refactors, migrations, and multi-file architectural changes. The gap between “AI suggested a snippet” and “AI shipped a PR” has basically closed.

3. Plugins and marketplaces.

The /plugin system turned Claude Code into a real ecosystem. You add a marketplace (/plugin marketplace add anthropics/skills), browse, install, and your agent gets new capabilities instantly. This is how Skills are actually distributed at scale today.

4. Routines.

The newest big feature: configure a Claude Code routine once, then trigger it on a schedule, via API, or in response to an event. Nightly dependency upgrades, auto-triage of new GitHub issues, on-merge changelog generation — all become one-time setup. Skills + Routines is the combo I’m most bullish on for 2026.

5. Real customers, real numbers.

Notion’s co-founder Simon Last said it best: “A big part of my job now is to keep as many instances of Claude Code busy as possible.” Ramp reported saving 1–2 days per ML model on Metaflow conversions. Intercom, Spotify, Shopify, Figma, Stubhub, and Asana have all gone public with Claude Code adoption. This isn’t early-adopter buzz anymore — it’s mainstream developer tooling.

6. Pricing finally makes sense.

Claude Code is bundled into Pro ($17–$20/mo), Max 5x ($100/mo), and Max 20x ($200/mo) plans. For the first time, “have the AI keep three parallel branches alive while I review the fourth” is economically sane.

Put it together and you get the real punchline: Claude Code is becoming the operating environment for AI-assisted engineering, and Skills are the package format for that environment.

How I Evaluated 100 Skills

Each Skill got scored on five axes:

Trigger precision — Does Claude pick it up at the right moment, and ignore it otherwise?
Determinism — Does it offload work to code where it should, instead of asking the model to “be careful”?
Token economy — Lean SKILL.md, with detail pushed into bundled files.
Reusability — Useful across multiple workflows, not a one-shot trick.
Safety posture — No surprising network calls, no opaque binaries, dependencies audit cleanly.

Anything that scored under 3/5 on more than two axes got cut. That eliminated about 70% of what I tried.

The Best 10 Claude Skills (Ranked)

1. PDF — The skill that made me a believer

Form filling, field extraction, and reliable text/table parsing without hallucination. The Skill ships with a Python script that reads PDFs and returns structured field metadata, so Claude executes the parser instead of imagining the contents. The forms.md companion file only loads when you’re actually filling a form. This is the canonical example of progressive disclosure done right.

Use it when: Anything PDF — extraction, form filling, batch redaction.

2. DOCX / PPTX / XLSX — Office, finally automated properly

The Office trio is the secret behind Claude’s document-creation features. They generate genuine .docx, .pptx, and .xlsx files (not Markdown pretending to be Word), preserve styles, and handle templates cleanly.

Killer combo: Use xlsx + pptx together to turn a CSV into a board-ready deck in one prompt.

3. skill-creator — The meta-skill

A Skill that helps you write Skills. It enforces the frontmatter contract, suggests good description wording (the part most people get wrong), and scaffolds bundled files. If you’re going to install one Skill before any other, install this.

Pro tip: Pair it with mcp-builder and you’ve got a self-bootstrapping agent toolkit.

4. mcp-builder — Bridge to the wider tool ecosystem

Generates Model Context Protocol (MCP) servers from a description. Skills + MCP is the combo Anthropic has clearly been building toward: Skills teach the workflow, MCP exposes the external tools. This Skill makes that pairing trivial.

5. webapp-testing — Playwright, but Claude drives

Spins up Playwright sessions, navigates flows, captures screenshots, and reports failures with structured output. I replaced an entire smoke-test script with “use the webapp-testing Skill on staging” and it worked first try. Wire it into a Claude Code Routine and you have nightly UI smoke-tests with zero CI YAML.

Caveat: Sandbox the browser. Always.

6. frontend-design — Designs that don’t look AI-generated

Encodes spacing, typography, and layout principles instead of vibes. The Skill nudges Claude to use semantic tokens, consistent scales, and accessible color contrast. Pairs beautifully with…

7. brand-guidelines — Your style guide as a Skill

Drop in your color palette, logo rules, voice and tone, and approved typography. Every artifact Claude generates afterward — slides, docs, web pages — comes back on-brand. This is the Skill enterprises have been quietly desperate for.

8. theme-factory — Design systems on demand

Generates cohesive themes (light/dark, semantic tokens, component variants) you can drop into Tailwind, CSS variables, or design tools. The output is structured JSON, not “here’s a vibe” — meaning it composes with code generators downstream.

9. internal-comms — The Slack-message ghostwriter you didn’t know you needed

Templates for announcements, status updates, incident comms, and exec summaries. The Skill teaches Claude your org’s tone — concise, no jargon, link-heavy, whatever. Saves me ~30 minutes a day on Slack alone.

10. slack-gif-creator — The unserious pick that earned its spot

Generates short, on-message animated GIFs for Slack reactions. Yes, it’s silly. Yes, it has driven measurable team morale gains. The Skill demonstrates how narrow a great Skill can be and still earn its install.

Honorable Mentions

algorithmic-art — Generative SVG/Canvas art with parameterized seeds. Great demo of code-execution Skills.
canvas-design — HTML5 canvas compositions for marketing assets.
doc-coauthoring — Multi-pass editing with diff-style suggestions; pairs with docx.
claude-api — Up-to-date reference for the Claude API itself, including Managed Agents, multiagent, and webhooks. Underrated.
web-artifacts-builder — Builds self-contained HTML artifacts (mini-apps, dashboards). Perfect for demos.
Notion Skills for Claude (partner) — Best partner Skill I’ve tested. Treats Notion like a first-class workspace, not an API surface.

Patterns I Saw in Every Great Skill

After 100 of these, the great ones rhyme:

A description that reads like a router rule.

Bad: “Helps with documents.”

Good: “Use when the user asks to extract form fields, fill, redact, or parse tables from a PDF file.”
Code where code belongs.

The model doesn’t sort lists, parse PDFs, or compute hashes. It calls a script. Cheaper, deterministic, repeatable.
Lean SKILL.md, fat reference.md.

The core file should fit on a phone screen. Push edge cases into bundled files Claude will only open when needed.
One Skill, one job.

Skills that try to do five things trigger at the wrong time and confuse the agent. Split them.
Examples > rules.

A Skill with three concrete worked examples beats a Skill with twenty bullet-pointed rules every time.

Patterns I Saw in Every Bad Skill

The 4,000-token SKILL.md that loads on every adjacent task.
Vague triggers like “use this for productivity tasks.” Productivity is not a category.
Self-reported metadata — Skills that claim to do things their bundled code can’t actually do.
Untrusted network calls baked into scripts with no documentation.
No examples. If I can’t guess the use case from the README, neither can the model.

A Word on Security (Read This Part)

Skills are powerful precisely because they let an agent execute code and follow instructions you didn’t write. That’s also exactly why they can be dangerous.

Before installing any Skill from a less-trusted source:

Read SKILL.md end-to-end. Look for instructions to fetch URLs, exfiltrate files, or call eval-style patterns.
Audit every script. Pin dependencies. Diff updates.
Sandbox execution. Containers, restricted file system access, network egress rules.
Treat partner badges as marketing, not assurance. Verify yourself.

Anthropic’s own guidance is blunt: install only from trusted sources, and audit anything else. That’s the right posture.

How to Try These Yourself

In Claude Code (recommended):

Install Claude Code first — one-liner from the docs:

irm https://claude.ai/install.ps1 | iex   # Windows
# or: curl -fsSL https://claude.ai/install.sh | sh   # macOS/Linux

Then wire up the official Skills marketplace:

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

Then just mention the Skill in a prompt:

“Use the PDF skill to extract the form fields from ./contracts/nda.pdf.”

From there you can drive Claude Code from the terminal, the IDE plugin, the desktop app (with parallel tasks and visual diffs), the web, iOS, or Slack — same Skills, same context.

As a Routine: Once a Skill-driven workflow proves itself, promote it to a Claude Code Routine so it runs on a schedule or in response to GitHub/webhook events. This is where Skills stop being a parlor trick and start replacing scripts.

In Claude.ai: The example Skills are available on paid plans — enable them in settings, then invoke by intent.

Via the Claude API: Upload custom Skills through the Skills API and reference them per request — ideal for embedding into your own product.

Author your own: Start from the template/ folder in anthropics/skills, run it through skill-creator, and iterate against real tasks. The fastest feedback loop is to author the Skill inside Claude Code itself — ask Claude to capture the steps it just took into a reusable SKILL.md.

Final Take: Skills + Claude Code Is the Combo to Watch

Tools (MCP) give agents capability. Skills give agents competence — the procedural knowledge to use those capabilities well, in your context, for your workflows. Claude Code gives both a home: a multi-surface, model-agnostic, plugin-enabled environment that’s gone from CLI curiosity to mainstream developer platform in under a year.

The Skills ecosystem is barely six months old and already feels like the format the industry has been quietly missing. Pair it with Claude Code’s Routines, parallel task management, and IDE/Slack/desktop reach, and you have something genuinely new: an agent that doesn’t just help you code, but learns the way your team works and quietly gets better at it every week.

If you’re building anything with Claude — or any agent that adopts the open standard — start with the ten Skills above, write your eleventh yourself, install Claude Code, and let your agent get genuinely good at the work you actually do.

About the Author

Suraj Khaitan — Gen AI Architect | Building scalable platforms and secure cloud-native systems

Connect on LinkedIn | Follow for more engineering and architecture write-ups

Which Claude Skill has changed your workflow the most? Drop your pick in the comments — I’m always hunting for the next great one.

DEV Community