Claude Code skills are folders of instructions, scripts, and resources that Claude Code discovers and activates on its own when your task matches - so your AI coding agent stops forgetting your stack, conventions, and workflow between sessions. You explain your preferences once; Claude applies them every time, without being reminded.
This guide covers what Claude Code skills are, the two kinds you'll use, how they auto-activate, where they live, and a battle-tested workflow for shipping real features with them.
The problem they solve
AI coding assistants forget everything between conversations. Without help, you re-explain your framework, your folder structure, your naming rules, and your review standards in every new session. A skill captures that knowledge once, so Claude Code follows your conventions consistently instead of defaulting to generic patterns that produce "AI slop."
The two kinds of Claude Code skills
- Capability uplift - skills that give Claude new abilities it didn't have: things like document generation, web app testing, or scraping. They extend what the agent can do.
- Encoded preference - skills that make Claude execute your way: frontend design rules, React best practices, your team's commit conventions. They constrain how the work gets done.
Most developers end up with a mix: a few capability skills for new powers, and several preference skills that enforce house style.
This is the "your review standards" example above made concrete: drop Yuri in and Claude checks every diff for bugs, security holes, and performance issues the way a senior reviewer would, every session.
View Yuri - Code Reviewer AI Skill →How they activate
You don't have to call a skill by name. Claude Code watches the conversation and activates a skill on its own when what you describe matches the skill's description. Mention reviewing code for vulnerabilities, and a security-review skill can fire automatically; ask for a component, and your React-standards skill kicks in. The description in each skill's SKILL.md is what controls this, which is why writing a sharp description matters more than anything else.
Where Claude Code skills live
Skills sit in a skills directory and load at startup:
-
Personal:
~/.claude/skills/<name>/- available across all your projects. -
Project:
.claude/skills/<name>/- committed to the repo so the whole team shares it.
Confirm a valid SKILL.md sits at the folder root, restart Claude Code, and run /skills to verify it loaded. Installing a packaged skill is often just unzip skill.zip -d ~/.claude/skills/.
Skills vs CLAUDE.md vs slash commands vs MCP servers
These four get confused constantly, and they solve different problems.
| Mechanism | What it is | When it loads | Best for |
|---|---|---|---|
| CLAUDE.md | Project memory file at repo root (or ~/.claude/CLAUDE.md) | Always, every turn | Facts Claude must never forget: stack, conventions, commands, do-not-touch areas |
| Skill | Folder with SKILL.md and optional support files | On demand, when the description matches | Expert procedures you need sometimes: reviews, audits, test plans, runbooks |
| Slash command | Markdown file in .claude/commands/, run with /name | Only when you type the command | Repeatable actions you trigger deliberately: /deploy-checklist, /write-pr |
| MCP server | External process exposing tools (databases, APIs, browsers) | Tools are available all session | Giving Claude the ability to do something, not the knowledge of how to do it well |
The short version: CLAUDE.md is always-on context, a skill is expertise that arrives when needed, a slash command is a button you press, and an MCP server is a set of hands. A good setup uses all four: MCP so Claude can act, a skill so it acts like a senior engineer.
One consequence: do not paste a 2,000 word review checklist into CLAUDE.md. It will be in context when you are renaming a variable. Make it a skill.
How to test a skill
Testing takes five minutes.
- Trigger test. Ask three requests that should fire the skill using different phrasing ("review this", "anything wrong with this function?", "check the PR"). If one misses, extend the description with that phrasing.
- Negative test. Ask something adjacent that should not fire it ("rename this variable"). If the skill loads anyway, the description is too broad.
- Output test. Compare the answer to what the SKILL.md promises. If the skill says "always list severity per finding" and the output has no severities, the instructions are being read but not followed; shorten and sharpen them.
- Explicit test. Name the skill directly. If that works and auto-discovery does not, the problem is 100 percent the description.
To see what a well built SKILL.md looks like before you write or buy one, 12 real SKILL.md files are explained line by line here.
A proven workflow: research → plan → implement
The most reliable way to ship a non-trivial feature with Claude Code is to separate thinking from typing:
-
Research. Ask Claude to map the relevant code first - existing patterns, dependencies, and extension points - into a short
research.md. No code yet. - Plan. Have Claude draft a plan with no implementation. Open it in your editor, annotate everywhere it's wrong, then send it back with: "address all notes, don't implement yet." Repeat until every decision is resolved.
- Implement. Only now let Claude write code, following the agreed plan.
A "feature-dev" preference skill can encode exactly this loop so it runs the same way every time - which is the whole point of skills: turning a good workflow into a repeatable one.
Instead of building the research → plan → implement loop from scratch, load a full bench of 5 coding agents into Claude and cover code review, bug fixing, docs, DevOps, and full-stack implementation in one pack.
View Coding Agent Bundle →What makes a Claude Code skill worth keeping
Good skills encode decisions, not information. A skill that explains what a REST API is wastes tokens on something the model already knows. A skill that says "in this repo every endpoint returns errors in the RFC 7807 shape, and the three places that do not are legacy and must not be copied" is worth keeping, because that is a decision the model cannot infer. The test: if a competent engineer joining your team would need to be told it, put it in a skill. If they would already know it, leave it out.
The categories that earn their place: code review with your house standard, testing conventions (framework, naming, what gets mocked), debugging and incident work (where the logs are, what to check before escalating), infrastructure and deployment (what is safe to run against production), and documentation in your house style.
What to skip:
- Language tutorials. A skill explaining Python syntax is explaining it to something that already knows Python.
- Catch-all skills. One skill covering frontend, backend, testing and deployment gets loaded constantly and is mostly irrelevant every time. Split it.
- Anything stale in a month. Version numbers, sprint goals, who is on call. Skills are for durable decisions.
- Vague descriptions. If the description does not say when the skill applies, it is either never selected or always selected.
Where to get Claude Code skills
You can write your own, or install ready-made ones. If you'd rather start from tested files, browse the Claude Code skills collection or the broader Tech & Development library. New to skills entirely? Start with how to install a Claude Skill or grab a free skill to see the structure.
Route 3: plugins
Skills can also ship inside a Claude Code plugin, installed with the plugin command rather than by moving files. Plugin skills are namespaced by their plugin, which keeps two skills with the same name from colliding.
The complete list of engineering skills worth installing
The useful skills for a CLI tool are the ones tied to a real engineering job title, because that is how the work is actually divided. Here is the full grouping we maintain for Claude Code, fourteen roles in total:
- Review and quality: Code Reviewer, Code Auditor, QA Engineer.
- Building: Full Stack Developer, Backend Developer, Frontend Developer, WordPress Developer.
- Infrastructure: DevOps Engineer, Cloud Engineer, Database Engineer.
- Design and direction: Software Architect, Tech Lead, Engineering Manager.
- Security: Cybersecurity Analyst.
Fifteen specific files we would start with, grouped by the job they do (prices are live at the time of writing):
Code quality
Yuri - Code Reviewer AI Skill ($19): Runs a structured review on any diff, file or PR with findings grouped by severity.
Soren - Code Auditor AI Skill ($19): Audits a whole codebase rather than a single change, looking for structural debt, dead code, inconsistent patterns and risk hotspots. Good for the first week on an inherited repository.
Albert - AI Code Review Agent ($39): A review agent tuned for a fast loop, with a focus on correctness and edge cases in the changed lines.
Mira - QA Engineer AI Skill ($19): Turns a feature description or a diff into test cases, edge cases and a test plan, then writes the tests in your framework. Useful when you ask Claude Code to "add tests" and want more than the happy path.
Delivery
Rami - DevOps Engineer AI Skill ($19): CI/CD pipelines, Dockerfiles, infrastructure as code and incident runbooks with the caution a real DevOps engineer brings. It asks about rollback before it writes a deploy step.
Max - Full Stack Developer AI Skill ($19): The generalist: scaffolds features across API, data layer and UI, and keeps the pieces consistent. If you only install one delivery skill, this is the one.
Lena - Frontend Developer AI Skill ($19): Component structure, state, accessibility and responsive behaviour for modern frontend stacks. Stops Claude Code from shipping a div soup with inline styles.
Niko - Backend Developer AI Skill ($19): API design, data modelling, error handling and performance for server code. Strong on the boring parts (validation, pagination, idempotency) that generic prompts skip.
Elena - Data Engineer AI Skill ($19): Pipelines, schemas, transformations and data quality checks. Use it when the task involves moving data rather than serving it.
Docs and process
Saoirse - Technical Writer AI Skill ($19): READMEs, PR descriptions, changelogs, architecture notes and onboarding docs in a consistent voice. Ask Claude Code to "write the PR description" with this installed and you get context, changes, testing and risks, not a one-liner.
Dorian - AI API Documentation Agent ($39): Generates and maintains API reference docs from the code: endpoints, parameters, examples, error responses. Point it at a router file and let it fill the gaps.
Verity - Build & Release Engineer AI Skill ($19): Versioning, release notes, commit message conventions, build configuration and reproducibility. The skill to install if your team argues about Conventional Commits.
Security
Anaya - Application Security Engineer AI Skill ($19): Reviews code for injection, auth flaws, secrets handling and dependency risk, and explains the fix rather than only the finding. Run it before a release, not after.
Selwyn - Web App Pentester AI Skill ($19): Thinks like an attacker against your own application: threat models, attack paths, and a prioritised list of what to test.
Cassandra - Vulnerability Management Engineer AI Skill ($19): Triage for scanner output and CVE lists: what matters, what is noise, what to patch first. Pair it with a dependency audit in your CI.
The whole set is in the Claude Code skills collection. Two of them below, so you can see what the descriptions look like in practice.
CI/CD pipelines with caching, staging gates and rollback triggers, multi-stage Docker builds, Kubernetes manifests and Helm charts, Terraform modules with remote state, and a monitoring stack with real alert thresholds instead of placeholders.
View Rami - DevOps Engineer AI Skill →
A one-line verdict, then findings ranked Critical to Medium across correctness, security, performance and maintainability. Every finding names the exact input or sequence that triggers it, and ships a corrected snippet.
View Soren - Code Auditor AI Skill →Free on GitHub or paid: what is the actual difference?
Both exist, and pretending otherwise would be dishonest.
Anthropic publishes example skills openly at github.com/anthropics/skills, and there are community repositories collecting more. They cost nothing, you can read every line, and for common cases they are a perfectly good starting point. If you only want to see the mechanism work once, start there.
What you tend to get for money is depth and consistency rather than a different mechanism. The gap usually shows up in four places:
- Trigger precision. Free skills are often written by someone solving their own problem, so the description assumes their setup. A description written for strangers has to be narrower and clearer.
- An example of finished output. A pack that ships a worked example is showing you the deliverable before you pay. A repository with a 40-line README is asking you to guess.
- Scope discipline. One skill covering backend, frontend, infrastructure and security gives four half-answers. Four role-shaped skills give four right ones.
- Maintenance. A repository last touched a year ago still references tools and versions from a year ago.
Do desktop Claude skills work in Claude Code?
Mostly, with one thing to know.
The skills you upload in the Claude app and the skills Claude Code reads from disk follow the same idea and the same file: Markdown with frontmatter. The packaging differs. The app takes a ZIP through its interface, while Claude Code reads a folder from the filesystem. They are also stored separately, so installing a skill in one place does not make it appear in the other. You install it twice if you want it in both.
Claude Code additionally supports frontmatter options that only apply to the CLI. A skill written for the app works in Claude Code; a skill that leans on CLI-only options will lose those extras if you upload it to the app, though the method itself still applies. Step-by-step instructions for the app route are in how to install a Claude skill.
If you'd rather skip writing your own SKILL.md for code review, Albert is the packaged version - a structured Senior Code Review Agent you load into Claude and use immediately, across any language or framework.
View Albert - AI Code Review Agent →Start with one skill
Pick the part of your workflow you re-explain most - your stack, your review rules, your commit style - and encode it. Browse the Claude Code skills collection for tested starting points, or read what Claude Skills are first.
In summary:
If you review code often, start with Yuri - Code Reviewer AI Skill ($19) or the ready-made Albert - AI Code Review Agent ($39); if you want your whole research → plan → implement loop covered, the Coding Agent Bundle ($99) gives you all 5 coding agents in one pack. All work in Claude, ChatGPT and any AI chat with no coding and a 30-day money-back guarantee.
