Skip to main content

Best Claude Code Skills to Use First: Picks by Workflow

Use /code-review to look for bugs, /simplify to clean up changed code, and an official skill when design, browser testing, or document work needs a more specific method. Install for the right scope, then check discovery and a real task result.

LaoZhang AI TeamPublishedUpdated 12 min read
On this page
Choosing bundled, official, and project-specific Claude Code skills for the work at hand

The best Claude Code skill to start with is the one that addresses your current bottleneck. For a code change, try the bundled /code-review to find correctness bugs or /simplify to clean up the implementation. For UI work, consider Anthropic's frontend-design; for repeatable browser checks, webapp-testing; for office files, document-skills. A custom project skill becomes useful when your team keeps explaining the same procedure.

You can start with a bundled skill immediately if it is available in your session. Install an additional package when its method matches a job you can name and evaluate. The recommendations below follow the official documentation and skill files checked on October 12, 2026; they are workflow choices, without a claimed performance ranking or hands-on benchmark.

Which Claude Code skills should you use first?

Your immediate jobFirst choiceWhat to look for in the result
Find bugs in a diff or pull requestBundled /code-reviewFindings that identify a concrete correctness issue and the affected code
Make changed code easier to maintainBundled /simplifyApplied cleanup that you can inspect in the diff
Build a UI with a deliberate visual directionOfficial frontend-designA design and implementation grounded in your audience, content, and brief
Check a local web app in the browserBundled /verify for a changed app; official webapp-testing for a reusable Playwright testing methodAn observed interaction and its outcome, with failures distinguished from unavailable prerequisites
Read or create PDF, Word, PowerPoint, or Excel filesOfficial document-skillsThe requested file or extraction, checked against the input and requirements
Work on Anthropic API codeBundled /claude-apiGuidance for the project's language and the actual integration task
Build an MCP serverOfficial mcp-builderA server implementation with usable tool schemas, error handling, and tests
Capture a repeated team procedureA project skill, optionally developed with skill-creatorThe same required steps and output across representative tasks

The bundled choices come from the current Claude Code command reference. The installable choices are documented in Anthropic's skills repository. The result column is a practical way to evaluate your own task, rather than a promise that invoking a skill guarantees it.

Start with bundled skills for everyday code work

Use /code-review for bugs and /simplify for cleanup

These skills address different questions. /code-review examines the current diff, or a specified pull request, branch, or path, for correctness bugs. /simplify reviews changed code for cleanup opportunities and applies fixes; its review does not look for correctness bugs. Depending on the model and effort level, code review can also cover cleanup, but cleanup alone is not a bug review. See the official command definitions.

For example, after changing a reimbursement form's validation, a review should determine whether the new logic rejects or accepts the wrong amounts. Cleanup should improve the code's structure and reuse without substituting for that correctness check.

At the Claude Code prompt, the basic invocations are:

/code-review
/simplify

Inspect the resulting findings and diff before treating the work as complete. The code-review flags also matter: --fix applies findings, --comment posts them to a pull request or merge request, and ultra starts a deep cloud review. Choose those actions deliberately; a local review request need not become a public comment.

Use /verify when the changed app needs to be observed

The bundled /verify builds and runs an app to observe whether a code change does what it should. /run launches and drives the app; /run-skill-generator can record a project's build and launch recipe for later use. /verify runs only when you invoke it. These are useful starting points when passing tests still leaves the user interaction uncertain. Official app verification documentation.

A project with a database, environment configuration, or a multistep startup may need a recorded recipe. A skill supplies a method; the runtime still needs the app's dependencies and an environment in which the relevant interaction can run. If a browser cannot start, that is a blocked browser check, not evidence that the form works.

Keep /debug, /loop, and /batch tied to their actual jobs

/debug enables session debug logging and investigates Claude Code issues through that log. If logging was off, it begins capturing from that point forward. It is not a general replacement for your application debugger.

/loop repeats a prompt while the session stays open. For instance, a developer could ask it to periodically check a deployment's status. It is useful for repetition within an open session, rather than a permanent background monitoring service.

/batch coordinates larger codebase changes through independent worktrees and subagents after presenting a plan. Its documented workflow includes publishing changes, so it deserves more consideration than a first experimental install. For Anthropic API work, start with /claude-api, which loads language-specific references and can also activate when code imports anthropic or @anthropic-ai/sdk. These behaviors are described in the current command reference.

The bundled set changes over time, and availability depends partly on the version, features, and environment. Use your session's command menu and the current reference to see what you can invoke.

Choose an official skill for the workflow you repeat

Workflow choices for browser testing, Anthropic API work, documents, and repeated team procedures

Frontend design: give it a real brief

Choose frontend-design when Claude needs a more deliberate approach to a new UI or an existing interface's visual direction. Its instructions emphasize the subject, audience, typography, palette, layout, and critique against the brief. It also includes responsive behavior, visible keyboard focus, and respect for reduced motion.

Supply enough information to make a design choice meaningful. A reimbursement form for employees has different priorities from a public conference page: the form needs clear field labels, reviewable totals, and useful validation feedback. Ask for those outcomes instead of “make it look premium.” You can then assess whether the implementation serves that brief.

The official repository offers this skill as part of its examples. Do not assume it is already bundled into every Claude Code installation just because a roundup says so; check /skills in your session.

Browser testing: pick webapp-testing for a repeatable test method

Choose webapp-testing when your recurring job is to interact with a local web application, inspect the rendered page, capture screenshots, or examine browser logs. The skill uses Python Playwright scripts and includes a helper for managing server startup.

This is a useful complement to /verify: the bundled skill addresses observing a code change, while webapp-testing supplies a specific browser-testing toolkit. For the reimbursement form, specify the starting state, the amounts entered, the expected validation message, and where you want screenshots saved. That produces a task whose outcome you can inspect.

Installing the instructions does not automatically install every browser, Python package, application dependency, or service the test needs. Have Claude distinguish a completed interaction from a missing prerequisite in its report.

Document work: install document-skills when files are the deliverable

Anthropic documents document-skills for PDF, DOCX, PPTX, and XLSX work. A concrete starter task could be extracting the form fields from an existing PDF and reporting each field's name and type. For generated documents, check the actual file rather than accepting a chat description of it. The repository README provides the install route and a PDF extraction example.

The repository describes its skills as demonstration and educational material and warns that behavior can differ across Claude products. Many examples use Apache 2.0 licensing, but the document skills are source-available, not open source. Check the relevant license when adapting them for redistribution. Official provenance helps identify the source; it does not establish that a file conversion or presentation meets your requirements.

MCP development and custom procedures: add them when the job exists

mcp-builder guides server development: understand the external API, design tools and schemas, implement authentication and pagination, and build and evaluate the result. Choose it when you are creating an integration. If you simply need access to an existing service, connecting a suitable server is a separate task; installing this guide does not create an authenticated connection.

skill-creator helps turn a defined task into a skill, try representative prompts, compare results, and improve the instructions and triggering description. It becomes worthwhile after you know the procedure you want to preserve. For a team that repeatedly checks release notes against actual changes, define the inputs, required checks, and expected output first. The skill should capture that process rather than invent a new one on every invocation.

Install official Anthropic skills in the right scope

The following are commands entered inside a Claude Code session, taken from the official repository README:

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

For document work, use this second command instead:

/plugin install document-skills@anthropic-agent-skills

anthropics/skills is the repository source; anthropic-agent-skills is the marketplace name used in those install commands. It differs from claude-plugins-official. Copy the marketplace suffix that belongs to the chosen package.

In current interactive Claude Code, /plugin install opens the plugin's details panel before installing. Review the components, then choose a scope:

ScopeChoose it whenWho gets the enabled plugin
UserYou want the workflow across your local projectsYou, on this machine
LocalYou want to try it in one repositoryYou, in this repository
ProjectThe team has agreed to share itCollaborators in the repository

Project settings enable a plugin for teammates but do not download it onto their computers. Each collaborator still needs to install it. For an individual first trial, local scope keeps the choice limited to the repository. Official installation and scope instructions.

The shell route is different: claude plugin install installs directly and defaults to user scope. To explicitly limit the example package to the current repository, the shell form is:

bash
claude plugin marketplace add anthropics/skills
claude plugin install example-skills@anthropic-agent-skills --scope local

The first command registers the marketplace; the second installs the example package at local scope. For documents, substitute document-skills@anthropic-agent-skills in the install command. The official plugin guide explains the shell route and its scope behavior.

Confirm the skill loaded, then check one useful result

Installation, discovery, and task success are separate checks. After installing, follow this short sequence:

  1. Read the install summary. “Plugin is now active” means no reload is needed. If it reports a server needing configuration, complete that setup before expecting its tools to work. If loading failed, open the Errors tab in /plugin.
  2. Apply pending changes when requested. Use /reload-plugins if the summary calls for it. If the reload warns that changing MCP tools would invalidate the prompt cache, understand that --force causes one uncached request before choosing it.
  3. Check discovery. Open /plugin → Installed to check scope and status, and /skills to find the skill and its source. Plugin skills use the form /<plugin>:<skill>; invoke the name shown by your session instead of guessing a short name.
  4. Run a bounded task. Give the input, expected outcome, and output location. For example: ask the discovered browser-testing skill to check that a negative reimbursement amount produces the specified message and save a screenshot.
  5. Inspect the output. A screenshot and the observed message support that interaction. A success statement with no completed browser action does not. For an edited UI, inspect the diff and running page; for document extraction, compare the reported fields with the input.

The activation messages, namespace, and discovery steps follow Anthropic's plugin installation guide. A single successful task tells you about that task; keep checking the outcomes that matter as you adopt the workflow.

When a project skill—or another mechanism—is the better choice

Project guidance, hooks, skills, subagents, and MCP serving different development needs

A project skill lives at .claude/skills/<skill-name>/SKILL.md and can include supporting scripts, references, and assets. Commit it when the team should share the procedure. A personal skill under ~/.claude/skills/ applies across local projects on that machine. The official location rules also cover nested repositories and plugin skills.

Write a description that says what the skill does and when it is relevant. By default, you can invoke it directly, and Claude can choose to use it when relevant. Use disable-model-invocation: true when you want to trigger the workflow yourself; user-invocable: false instead hides direct invocation. These controls address invocation timing. An allowed-tools entry grants permission for listed tools during the invoking turn; it is not a tool whitelist that makes all other actions impossible. Invocation controls, tool permission behavior.

Keep broad project facts and conventions in project guidance such as CLAUDE.md. Use a hook for a handler that must run at a configured event, an external integration for missing tools or data access, and a subagent when the work needs delegated execution. A skill can teach a procedure and work alongside those mechanisms. For the fuller comparison and a minimal custom skill or hook, read Claude Code Hooks vs Skills vs Slash Commands.

Fix missing skills, unwanted activation, or the wrong version

SymptomFirst actionWhat it helps distinguish
A plugin skill is missingCheck Installed and Errors in /plugin, then the install summaryDisabled, pending reload, or failed loading
The bare name invokes something elseFind the skill's source in /skills and use its displayed namespacePlugin naming versus a same-name local skill
A new project skill is missingCheck its location and use /reload-skills if the top-level skills directory was created after startupDiscovery failure versus an invocation problem
Direct invocation works, but natural requests do notCheck the description, YAML, and disable-model-invocationMissing trigger metadata or deliberately manual-only use
The skill activates on unrelated workNarrow its description or make it manual-onlyAn overly broad trigger
A remembered bundled command behaves differentlyCheck /skills for a custom skill with the same nameA local replacement for the bundled command

Claude Code normally watches existing skill directories, so every skill edit does not require restarting. For plugin hooks, MCP servers, agents, or output styles, use plugin reload rather than treating them as a plain SKILL.md edit. Malformed YAML can leave the body available but metadata empty; claude plugin validate can check skill directories on v2.1.233 or later. Skill editing, troubleshooting.

If many skills are available but Claude misses relevant ones, inspect /context: the skill listing has a budget and can drop descriptions while retaining names. /skill-doctor, where available, helps identify context cost and usage; it requires v2.1.252 or later and feature-flag fetching. This is a reason to keep the useful set understandable, not a rule that every developer must install a fixed number of skills.

To stop an experiment, disable its plugin from Installed and apply the pending change. For removal, target the scope you actually used. For example, the shell command claude plugin uninstall example-skills@anthropic-agent-skills --scope local removes the local installation. When a project enables a plugin, the interactive uninstall offers “Disable for me” and “Uninstall for everyone”; choose according to your intended effect. Disabling or uninstalling does not undo files edited or actions taken earlier. Official plugin management.

FAQ

Do I need to install skills before using Claude Code effectively?

No. Bundled skills already cover tasks such as code review, cleanup, app verification, and Anthropic API guidance. Choose from what is available in your session before adding a package. The current command reference describes each one's purpose and conditions.

Which official skill should I install first for web development?

Choose frontend-design if the bottleneck is visual direction and UI construction. Choose webapp-testing if the bottleneck is repeatable local browser interaction. For observing a specific app change, try bundled /verify first. These choices follow the design skill, testing skill, and app verification documentation; none establishes a universal winner.

Is /simplify a code review?

It reviews changed code for cleanup and applies fixes, but does not look for correctness bugs. Use /code-review for bug finding. The distinction is explicit in the official command reference.

Are Anthropic's official skills all open source?

No. The repository README says many skills use Apache 2.0, while the document skills are source-available rather than open source. Check each skill's license before redistributing or adapting it.

Does a skill installed locally appear in Cowork or every cloud session?

A plugin installed on your machine does not automatically become a plugin enabled for your claude.ai account. Personal filesystem skills also do not load in Cowork or cloud sessions simply because they exist under your local home directory. Account-synced plugins and skills have separate sign-in and availability conditions. See plugin sync and skill locations.

Sources8

External pages this guide links to, in the order they appear. Last updated Oct 12, 2026.

  1. 1.Claude Code command referencecode.claude.com/docs/en/commands
  2. 2.Anthropic's skills repositorygithub.com/anthropics/skills
  3. 3.Official app verification documentationcode.claude.com/docs/en/skills
  4. 4.frontend-designgithub.com/anthropics/skills/blob/main/skills/frontend-design/SKILL.md
  5. 5.webapp-testinggithub.com/anthropics/skills/blob/main/skills/webapp-testing/SKILL.md
  6. 6.mcp-buildergithub.com/anthropics/skills/blob/main/skills/mcp-builder/SKILL.md
  7. 7.skill-creatorgithub.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md
  8. 8.Official installation and scope instructionscode.claude.com/docs/en/discover-plugins
More in Claude Code