10 Claude Skills. The Complete Map.

A skill is a folder Claude reads on demand. Not a prompt. Not an agent. Not a plugin. A markdown file plus a few supporting files that live in a specific directory, that Claude loads when the task calls for it, and that changes how Claude behaves for the duration of that task.

The article you are reading is the map of the ten skills that matter in July 2026. What each one does. When to reach for it. What breaks without it. Which one to install first. And the meta-skill that compounds the other nine.

Six Parts. Part 4 is the filter that picks your first three. Read that one twice if you read nothing else.

PART 1. WHAT A SKILL ACTUALLY IS

Three things a skill is not.

It is not a prompt. A prompt lives inside one conversation. A skill lives in a folder and gets pulled into any conversation on demand. The prompt burns. The skill accumulates.

It is not an agent. An agent is a loop with a goal, a verifier, and a stop condition. A skill is a capability layer the agent uses inside its loop. The two work together. They are not the same.

It is not a plugin. A plugin is code executed by the runtime. A skill is instructions and examples read by the model. The plugin runs. The skill teaches.

What a skill actually is: a small folder with a markdown file at the root that explains one specific way of doing one specific job, plus any supporting templates or code Claude needs when it does that job. When Claude sees a task that matches the skill's declared purpose, it reads the folder and follows what is inside.

The compounding advantage is that ten skills you built once and tested well replace a thousand prompts you had to remember and rewrite every session.

PART 2. THE 10 SKILLS

The ten skills below cover the surface area most builders on Claude actually need. They were built by ten different authors across 2025 and 2026. Each one is a folder you install, not a service you subscribe to.

1. Frontend Design. Forces Claude to generate frontend code with a defined aesthetic instead of the default cream-and-purple gradient template every model defaults to. Declares Brutalism, Minimalism, Retro-Futurism, or a custom named movement, then generates the components consistent with that choice.

What breaks without it: every generated landing page looks like the same 2024 Stripe clone. What it gives you: sites and dashboards that look intentional.

2. Algorithmic Art. Generative art via p5.js. Produces interactive HTML artifacts with sliders, seed inputs, and next/random buttons. Parameters change in real time. Variations get saved. The skill is the practical entry point into NFT collection generation, generative print art, and hero-image variation for landing pages.

What breaks without it: Claude produces one static image per prompt. You want a family.

3. Systematic Debugging. A four-phase bug-hunting process built on Root Cause Analysis. Instead of "let's try this, let's try that," the skill imposes a structural breakdown with explicit component-boundary logging. Reported outcome: bugs fixed in 15 to 30 minutes instead of 2 to 3 hours of chaotic guessing.

What breaks without it: you and Claude spend an afternoon trading hypotheses. What it gives you: a documented diagnostic trail.

4. Canvas Design. Creates posters, book covers, one-pagers, and static visuals in PNG and PDF. First defines a design philosophy (a named movement with its own principles), then executes every output against that philosophy. Not generic AI images. Design outputs with logic behind them.

What breaks without it: Claude produces stock-photo-quality decorative fluff. What it gives you: covers you would show a client.

5. Theme Factory. Ten ready-made professional themes (colors, typography, spacing scales) applicable to any artifact with a single command. Slides, documents, landings, reports, dashboards. Or generate a new theme on the fly and apply it across every artifact in the same session. Saves the 20-minute session tax of manually setting up styling before you can start building.

What breaks without it: every artifact looks like a hackathon demo. What it gives you: consistency across everything you ship in a week.

6. Web Artifacts Builder. For complex web artifacts inside claude.ai. Scaffolds React 18, TypeScript, Tailwind, and 40+ shadcn/ui components, then bundles the whole thing into a single HTML file with functional routing and state. If your reflex is to build a real app instead of a "hello world," this is the skill.

What breaks without it: your artifact is a single view with no state. What it gives you: apps that feel like apps.

7. Superpowers. An entire framework for agentic coding by Jesse Vincent. Includes 20+ sub-skills covering TDD, brainstorming, planning, code review, git worktree management, and subagent-driven development. Claude working under this framework runs autonomously for longer stretches without human intervention veering it off-plan. This is the tier that turns a coding session into an operation.

What breaks without it: you interrupt Claude every fifteen minutes to redirect. What it gives you: an evening of hands-off shipping.

8. File Search. From the MassGen project. Fast codebase search using ripgrep for text and ast-grep for AST-level syntax queries. When Claude is dropped into an unfamiliar repo, or you ask it to find every place a function should be refactored, this skill saves hours over the default grep and find calls.

What breaks without it: Claude reads whole files to answer questions two search calls would resolve. What it gives you: precision at repo scale.

9. Context Optimization. A collection of context-engineering skills covering compaction, tool-output masking, KV-cache reuse, and multi-agent context partitioning. If your Claude sessions run out of context after 30 minutes of real work, the issue is not the model. It is the absence of these skills.

What breaks without it: session dies at message 25 instead of message 400. What it gives you: full-day sessions without a context wall.

10. Skill Creator. The meta-skill. Teaches Claude how to create its own new skills, with evaluations, benchmarks, and tests baked in. Once installed, every new problem you and Claude solve becomes a new skill in your library. The other nine came from other builders. This one lets you produce your own.

What breaks without it: your library caps at the ten skills other people wrote. What it gives you: a library that grows every time you build something new.

PART 3. HOW TO INSTALL AND ORGANIZE

The install is a folder in a specific location. On Claude Code, skills live in a directory Claude scans at start-up. In Claude Cowork, skills load per project. The details vary by client. The principle is the same: a folder with a markdown file at the root, optional supporting files, and a name that matches the skill's declared purpose.

Recommended directory structure:

/claude-skills/
   frontend-design/
      SKILL.md
      references/
   algorithmic-art/
      SKILL.md
      examples/
   systematic-debugging/
      SKILL.md
      templates/
   ...

Two operational rules that separate a working install from a broken one.

First rule: do not load all ten skills at once. The context cost of ten skill files sitting in the system prompt is the entire reason your session died at message 25. Load skills on demand based on the current task. Well-written skills declare their trigger conditions, so the client handles this automatically. Badly written ones bleed into every conversation.

Second rule: name skills by outcome, not by tool. "systematic-debugging" is the outcome. "root-cause-analysis-framework" is the tool. When Claude sees a bug in the conversation, it reaches for the skill named after the outcome, not the tool.

PART 4. WHICH SKILL TO INSTALL FIRST

If you install all ten and try to use all ten, you will burn a weekend on setup and end up using two of them. The order matters. The order below is based on which skills reward the earliest installation.

If you are a frontend developer:

1. Frontend Design      (immediate output quality lift)
2. Web Artifacts Builder (working apps, not toys)
3. Theme Factory        (consistent styling across projects)

If you are a content creator or designer:

1. Canvas Design        (covers, one-pagers, static visuals)
2. Algorithmic Art      (variations for hero images and NFTs)
3. Theme Factory        (consistency across posts)

If you are a backend or systems developer:

1. Systematic Debugging (single largest hourly rate improvement)
2. File Search          (repo comprehension at scale)
3. Context Optimization (sessions that survive real work)

If you are running agentic workflows:

1. Superpowers          (the agentic tier)
2. Context Optimization (avoids the wall at long horizons)
3. Skill Creator        (extends the library from what you build)

The two skills every reader should install regardless:

Context Optimization  (universal: every session benefits)
Skill Creator         (the multiplier: extends the library)

Install the two universal skills first. Then install the three that match your work. Then install additional skills as new problems demand them. Do not install all ten on day one.

PART 5. WHAT BREAKS AT 10+ SKILLS

The library that solves problems at 5 skills starts creating them at 20.

Three failure modes to watch.

Context pollution. Every skill loaded into the system prompt costs tokens on every call. Ten skills at 800 tokens each is 8,000 tokens of overhead per message. If your sessions are getting cut off earlier as your library grows, this is the reason. Prune the always-loaded set. Move rarely-used skills to on-demand loading.

Skill collision. Two skills that both claim jurisdiction over the same task cause Claude to load both, follow both, and produce output that satisfies neither. Common example: two "design" skills, one for canvas, one for frontend, both getting pulled for a landing-page mockup. Solve by explicit disambiguation in each skill's declared purpose.

Cost multiplication. A conversation that loads five skills does not cost five times a plain conversation. It costs seven to twelve times, because the skills often reference each other's outputs. Model your cost per accepted output as your library grows. When the number climbs, audit which skills are pulling more than they contribute.

The rebuild trigger. When your library hits fifteen or more skills, take a weekend to audit. Consolidate overlapping skills. Delete skills you have not touched in three months. Rewrite the ones you use daily using what you learned from the ones you deleted. Every library gets one rebuild per year.

PART 6. THE META-SKILL

The nine skills above came from other builders. The tenth, Skill Creator, is what lets you produce the eleventh, twelfth, and hundredth.

Skill Creator is a meta-instruction set. It teaches Claude the shape of a good skill: the declared purpose, the trigger conditions, the required supporting files, the built-in evaluations that verify the skill still works after edits. Every problem you solve manually one afternoon becomes a Claude-generated skill in the evening.

The compounding effect over twelve months is the real product.

Month 1:  10 installed skills, 0 you wrote
Month 3:  15 skills, 5 you wrote
Month 6:  25 skills, 15 you wrote
Month 12: 40 to 60 skills, most of them yours
Month 24: a personal library that solves problems you never 
          have to re-explain to a model again

The person with the newest model in month 24 is not at the same starting line as the person with a 24-month skill library. The library is the moat.

Everyone talks about which model to use. The compound advantage is which skills you have already built.

That's the entire library.

1.  Frontend Design       github.com/anthropics/skills/tree/main/skills/frontend-design
2.  Algorithmic Art       github.com/anthropics/skills/tree/main/skills/algorithmic-art
3.  Systematic Debugging  github.com/obra/superpowers
4.  Canvas Design         github.com/anthropics/skills/tree/main/skills/canvas-design
5.  Theme Factory         github.com/anthropics/skills/tree/main/skills/theme-factory
6.  Web Artifacts Builder github.com/anthropics/skills/tree/main/skills/web-artifacts-builder
7.  Superpowers           github.com/obra/superpowers
8.  File Search (MassGen) github.com/massgen/massgen
9.  Context Optimization  github.com/muratcankoylan/agent-skills-for-context-engineering
10. Skill Creator         github.com/anthropics/skills/tree/main/skills/skill-creator

Ten skills. Six Parts. A filter that picks your first three. An install discipline that prevents context pollution at scale. A meta-skill that compounds every subsequent skill you build.

The skill library is the actual product Claude ships in 2026. The model is the runtime. The skills are the software you install on it. Everyone downloads the runtime. The people who install the software win the year.

→ Install Skill Creator and Context Optimization this week

→ Add the three that match your work by Friday

→ Rebuild the library every quarter to keep the ratio of yours to theirs climbing