Best AI skills in 2026: the Claude Skills and agent skills worth installing
Skills are the fastest-moving layer in AI tooling right now. Here are the ones that change how an agent works, not just what it knows.
A skill is a set of instructions an agent loads when it needs them — a way of encoding *how* to do something rather than hoping the model remembers. In 2026 this became the highest-leverage thing you can add to a coding agent, and the ecosystem now runs to tens of thousands of published skills. Most are noise. These are the ones worth your time.
Skill, MCP server, or plugin?
They solve different problems and people conflate them constantly. A skill teaches the model a method — how to review code, how to structure a diagram. An MCP server gives the model a capability it lacks — read this database, drive this browser. A plugin usually bundles both. If your agent knows what to do but does it badly, you want a skill. If it cannot reach something at all, you want an MCP server.
Where skills actually come from
There is no single official store, which is why finding good ones is hard. Three sources cover almost everything worth having: Anthropic's own anthropics/skills repository, the plugin marketplace built into Claude Code, and a handful of community collections that have earned real adoption. Everything below comes from one of those.
The engineering skills
Superpowers is the most ambitious of them — a full brainstorm-to-merge methodology that walks an agent through design spec, plan, subagent-driven build, review and merge. It is opinionated in a way that either fits how you work or does not, but it is the closest thing to a complete development process expressed as a skill.
Agent Skills (Addy Osmani's collection) is the pragmatic counterpart: testing discipline, performance work, review standards, refactoring patterns. Where Superpowers restructures your whole workflow, this raises the floor on the work an agent already does.
Andrej Karpathy behavioural skills encodes four rules distilled from Karpathy's observations about where LLMs go wrong writing software: think before coding, keep implementations simple, make surgical changes, verify the success criteria. It is the shortest thing on this list and probably the highest ratio of benefit to effort.
The best skills do not make the model smarter. They stop it doing the specific stupid thing it keeps doing.
The official Anthropic set
skill-creator is where to start if you plan to write your own — it scaffolds the structure and teaches the format by doing. mcp-builder turns a description of the tools you want into a runnable MCP server. webapp-testing and frontend-design are the two that most visibly change output quality: the first makes an agent actually verify a UI works, the second hands it a design philosophy so generated interfaces look considered rather than default.
The document skills — docx, pdf, pptx, xlsx — are unglamorous and quietly essential if you generate real deliverables. claude-api is worth adding to any project that calls Claude programmatically, because it keeps parameter shapes current instead of remembered.
Design and communication
Design engineering skills covers animation, interaction detail and interface craft — the judgment that separates a working UI from a considered one. Diagram design ships 29 editorial diagram types as self-contained HTML and SVG, which matters because the default failure mode is every explanation becoming the same flowchart. Theme factory generates coherent colour, type and spacing systems so generated interfaces share one visual language.
Security, research and the specialists
Trail of Bits Security Skills brings audit-grade workflows — CodeQL and Semgrep static analysis, variant analysis, fix verification — from the firm behind Slither and Echidna. Reverse skill routes an agent through reverse engineering and penetration testing methodology; for authorised work only, and worth saying so plainly. Book to skill converts a technical book PDF into a skill, which turns the book you meant to read into something the agent references while working.
Skills are instructions you are trusting
A skill is prompt text and sometimes executable tooling, usually written by a stranger. Read one before you install it, the same way you would read a shell script from a blog post. Popularity is a signal about usefulness, not about safety.
How to choose
- Does it solve a problem you actually hit? Installing skills speculatively is how you end up with a bloated context and slower runs.
- Does it reduce token overhead or add it? Every loaded skill costs input tokens on every turn. A skill has to earn that.
- Is it maintained? Check the last commit. A skill written against last year's agent behaviour can actively mislead.
- Start with three. Add one at a time and notice whether the output changed. If you cannot tell, remove it.
The skills layer is moving faster than any other part of AI tooling, and most of what is published this month will be irrelevant by the next. We keep the full skills directory current rather than letting a list rot — it is the same catalog this piece is drawn from.
Find these on the Radar
Every tool here lives on Kapyn Radar. Save the ones that fit into a Loadout and find them again.
