Dokumentationerreichbar
Anthropic, Extend Claude with skills
code.claude.com (externe Seite)
Die Herstellerdokumentation zu Skills, den abrufbaren Arbeitsanweisungen neben dem Modell. Sie liefert das Kostenargument dieses Kapitels: Der Rumpf eines Skills lädt erst beim Gebrauch, umfangreiches Nachschlagematerial kostet also bis dahin fast nichts, und sobald er geladen ist, bleibt er über die folgenden Runden stehen. Dazu die Steuerung darüber, wer ihn aufrufen darf, und der Maßstab für die Wirkung: Dass ein Skill anspringt, heißt nur, dass er gefunden wurde.
geprüft 10.11.2026
Worauf sich diese Seite beruft, wörtlich, abgerufen am 04.08.2026:
Create a skill when you keep pasting the same instructions, checklist, or multi-step procedure into chat, or when a section of CLAUDE.md has grown into a procedure rather than a fact.
bestätigt 10.11.2026Unlike CLAUDE.md content, a skill's body loads only when it's used, so long reference material costs almost nothing until you need it.
bestätigt 10.11.2026Once a skill loads, its content stays in context across turns
bestätigt 10.11.2026You don't want Claude deciding to deploy because your code looks ready.
bestätigt 10.11.2026Seeing a skill trigger tells you Claude found it, not that it did what you intended.
bestätigt 10.11.2026A fresh session matters because leftover context from authoring the skill will mask gaps in the written instructions.
bestätigt 10.11.2026The budget scales at 1% of the model's context window.
bestätigt 10.11.2026When the listing overflows, Claude Code drops descriptions starting with the skills you invoke least, so the skills you use most keep their full text.
bestätigt 10.11.2026When the conversation is summarized to free context, Claude Code re-attaches the most recent invocation of each skill after the summary, keeping the first 5,000 tokens of each.
bestätigt 10.11.2026Claude Code runs a hook every time its event occurs, such as before each file edit, whether or not Claude is following the skill.
bestätigt 10.11.2026Claude reports its findings first and asks for confirmation before changing anything.
bestätigt 10.11.2026Despite the name, a skill with context: fork doesn’t run in a fork of the current conversation
bestätigt 10.11.2026The subagent doesn’t see your conversation history, so the skill’s instructions have to stand on their own.
bestätigt 10.11.2026Claude uses skills when relevant, or you can invoke one directly with
bestätigt 10.11.2026Every skill needs a SKILL.md file with two parts: YAML frontmatter between --- markers that tells Claude when to use the skill, and markdown content with the instructions Claude follows when the skill runs.
bestätigt 10.11.2026The directory name, or the frontmatter name when you set one, becomes the command you type, and the description helps Claude decide when to load the skill automatically.
bestätigt 10.11.2026Keep SKILL.md under 500 lines.
bestätigt 10.11.2026Custom commands have been merged into skills.
bestätigt 10.11.2026both create /
bestätigt 10.11.2026deploy and work the same way. Your existing .claude/
bestätigt 10.11.2026commands/ files keep working. a Markdown file in .claude/
bestätigt 10.11.2026commands/ is the older format and still works. Skills add optional features: a directory for supporting files
bestätigt 10.11.2026By default, both you and Claude can invoke any skill.
bestätigt 10.11.2026After plain text, the name gives Claude permission to run the skill but doesn’t run it
bestätigt 10.11.2026To write about the skill without permitting a run, leave off the slash.
bestätigt 10.11.2026If Claude tries anyway, Claude Code blocks the call and instructs it not to reproduce the deploy steps another way
bestätigt 10.11.2026Where you save a skill decides which sessions load it.
bestätigt 10.11.2026Enterprise over personal, and personal over project.
bestätigt 10.11.2026Your skill replaces the bundled command, but not its aliases.
bestätigt 10.11.2026don’t read ~/.claude/
bestätigt 10.11.2026skills/ on your machine. Bundled skills are prompt-based: they give Claude detailed instructions and let it orchestrate the work using its tools.
bestätigt 10.11.2026Most built-in commands instead execute fixed logic directly.
bestätigt 10.11.2026which keeps you in control of when these longer-running checks spend time and tokens.
steht dort nicht mehrClaude Code runs the command and replaces the line with its output before Claude sees the skill content
bestätigt 10.11.2026the rendered SKILL.md content enters the conversation as a single message and stays there across later turns.
bestätigt 10.11.2026Claude Code does not re-read the skill file on later turns
bestätigt 10.11.2026All fields are optional. Only description is recommended so Claude knows when to use the skill.
bestätigt 10.11.2026Claude Code ignores a field it doesn’t recognize without reporting an error.
bestätigt 10.11.2026the combined description and when_to_use text is truncated at 1,536 characters in the skill listing to reduce context usage.
bestätigt 10.11.2026When set, Claude loads the skill automatically only when working with files matching the patterns.
bestätigt 10.11.2026Indexed arguments use shell-style quoting, so wrap multi-word values in quotes to pass them as a single argument.
bestätigt 10.11.2026If you invoke a skill with arguments but no placeholder in the skill’s content receives one, Claude Code appends
bestätigt 10.11.2026you can stack several skills at the start of one message.
bestätigt 10.11.2026Through the Agent SDK or claude -p, Claude Code loads only the first skill, with the rest of the line as its arguments.
bestätigt 10.11.2026doesn’t attach the files that @ references name the way it does for a local skill
bestätigt 10.11.2026include ultrathink anywhere in the skill content.
bestätigt 10.11.2026Claude never sees the skill content for that invocation.
bestätigt 10.11.2026With the default bash shell, any non-zero exit code counts as a failure.
bestätigt 10.11.2026Claude Code treats exit code 1 from search and comparison commands as a normal result and injects their output.
bestätigt 10.11.2026Injected commands never prompt for permission while the skill renders.
bestätigt 10.11.2026A command a deny rule matches aborts the invocation
bestätigt 10.11.2026when a command’s permission check returns anything other than allow, Claude Code aborts the invocation with the same error.
bestätigt 10.11.2026In auto mode, a command that would otherwise need your approval doesn’t abort the invocation.
bestätigt 10.11.2026Bundled and managed skills are not affected.
bestätigt 10.11.2026The directory containing the skill’s SKILL.md file.
bestätigt 10.11.2026Using the same variable in both places lets a skill run a bundled script without a permission prompt.
bestätigt 10.11.2026Claude Code starts a new subagent of the type set in the agent field and gives it the skill content as its prompt.
bestätigt 10.11.2026only makes sense for skills with explicit instructions.
bestätigt 10.11.2026The built-in Explore and Plan agents start without CLAUDE.md and git status to keep their context small
bestätigt 10.11.2026you keep working while it runs, and its result arrives in your conversation when it completes.
bestätigt 10.11.2026in the frontmatter to instead wait for the result in the turn that invoked the skill.
bestätigt 10.11.2026A backgrounded fork also runs with the narrower tool set that applies to background subagents
bestätigt 10.11.2026applies its edits outside your session’s checkpoints.
bestätigt 10.11.2026/
bestätigt 10.11.2026rewind restores only the files Claude had already edited in your conversation. The skill’s edits to other files stay, so use git to revert them.
bestätigt 10.11.2026The override applies for the rest of the current turn and isn’t saved to settings.
bestätigt 10.11.2026Overrides the session effort level, but not the CLAUDE_CODE_EFFORT_LEVEL environment variable.
bestätigt 10.11.2026The allowed-tools field grants permission for the listed tools during the turn that invokes the skill
bestätigt 10.11.2026The grant clears when you send your next message.
bestätigt 10.11.2026It does not restrict which tools are available
bestätigt 10.11.2026Workspace trust doesn’t gate this field.
bestätigt 10.11.2026A skill can grant itself broad tool access, so review the allowed-tools of skills checked into a repository before you run Claude Code there.
bestätigt 10.11.2026Claude Code ignores allowed-tools in project and personal skills
bestätigt 10.11.2026Permission syntax: Skill(name) for exact match, Skill(name *) for prefix match with any arguments.
bestätigt 10.11.2026The skill
bestätigt 10.11.2026Overrides setting controls skill visibility from your settings instead of the skill’s own frontmatter. highlight a skill and press Space to cycle states, then Esc to save to
bestätigt 10.11.2026Manage those through /
bestätigt 10.11.2026plugin instead. Claude Code picks up the change within the current session, without a restart.
bestätigt 10.11.2026Live change detection covers SKILL.md text only.
bestätigt 10.11.2026If you create a top-level skills directory that didn’t exist when the session started, run
bestätigt 10.11.2026Run /
bestätigt 10.11.2026skill-doctor to see what each of your skills costs and how often it gets used, so you can decide which ones to turn off. requires Claude Code v2.1.252 or later and isn’t available in sessions that skip
bestätigt 10.11.2026Claude Code loads project skills from .claude/
bestätigt 10.11.2026skills/ in the directory where you start it and in every parent directory up to the repository root Skills in a .claude/
bestätigt 10.11.2026skills/ directory below where you started don’t load at startup. They load the first time Claude reads or edits a file in that subdirectory and stay available for the rest of the session.
bestätigt 10.11.2026Until then they don’t appear in the / menu and you can’t invoke them by name.
bestätigt 10.11.2026grants file access only and loads none of these.
bestätigt 10.11.2026When the session starts, Claude Code downloads your account’s skills into ~/.claude/
bestätigt 10.11.2026skills/ synced/ in the background Claude Code syncs only in a session that signs in with your claude.ai account
bestätigt 10.11.2026Claude Code downloads synced skills and never uploads them.
bestätigt 10.11.2026To stop syncing on a machine, set syncClaudeAiSkills to false in your user settings.
bestätigt 10.11.2026skill loads Claude API and Managed Agents reference material for your project’s language.
bestätigt 10.11.2026Claude also activates it automatically when your code imports
bestätigt 10.11.2026Claude Code extends the standard with additional features like
bestätigt 10.11.2026Outside Claude Code, you can use only the fields in the Agent Skills spec
bestätigt 10.11.2026If you include any field the spec doesn’t allow, packaging or upload fails with a hard error instead of ignoring the field
bestätigt 10.11.2026don’t function in claude.ai chat or through the API.
bestätigt 10.11.2026When a session starts with a skill named verify or simplify in place, Claude Code’s commit instructions tell Claude to run it right before each commit, except for changes to docs or tests.
bestätigt 10.11.2026The directory name also invokes the skill.
bestätigt 10.11.2026the bundled alias /
bestätigt 10.11.2026review never runs your skill In a local terminal session, your skill replaces the built-in command, but not its aliases.
bestätigt 10.11.2026The invocation still aborts in a forked skill that sets agent
bestätigt 10.11.2026a later sync can overwrite or remove it.
bestätigt 10.11.2026Build and run your app to confirm a code change does what it should, without falling back to tests or type checks
bestätigt 10.11.2026That inference gets unreliable for projects that need anything beyond a standard launch
bestätigt 10.11.2026/
bestätigt 10.11.2026run-skill-generator records the recipe instead. At the repo root, the recorded skill replaces the bundled /
bestätigt 10.11.2026verify Claude edits the recorded file only when it steered a run wrong
bestätigt 10.11.2026The recipe that /
bestätigt 10.11.2026verify records at your repo root is a project skill, so it counts. Set to true to prevent Claude from automatically loading this skill. Use for workflows you want to trigger manually with
bestätigt 10.11.2026Also prevents the skill from being preloaded into subagents and from running when a session-scoped scheduled task fires with the skill as its prompt.
bestätigt 10.11.2026To remove tools from Claude’s available pool while a skill is active, list them in disallowed-tools in the skill’s frontmatter.
bestätigt 10.11.2026The restriction clears when you send your next message.
bestätigt 10.11.2026skills your organization provides there, and Anthropic’s built-in skills such as
bestätigt 10.11.2026That hook applies from the time the skill is invoked until the session ends.
bestätigt 10.11.2026invoke the skill again to restore its full content
bestätigt 10.11.2026Run /
bestätigt 10.11.2026doctor for an estimate of the listing’s context cost and its biggest contributors. setup checkup stays typable when disableBundledSkills is on.
bestätigt 10.11.2026Claude Code accepts the field but doesn’t act on it.
bestätigt 10.11.2026Claude Code searches parent directories only up to the worktree root.
bestätigt 10.11.2026gets a prefix when its folder name clashes with an enterprise, personal, or project skill
bestätigt 10.11.2026The usage check requires Claude Code v2.1.305 or later, and /
bestätigt 10.11.2026doctor leaves it out while Remote Control is connected. last 7 days
bestätigt 10.11.2026use square brackets or argument-hint
bestätigt 10.11.2026A custom command or skill doesn’t load when Claude Code refuses its directory or file name.
bestätigt 10.11.2026a personal or project skill whose frontmatter declares hooks loses its allowed-tools pre-approval
bestätigt 10.11.2026because its hooks don’t run under that setting
bestätigt 10.11.2026To keep them, move the skill into a plugin that you install from a marketplace.
bestätigt 10.11.2026don’t name a skill folder synced, in any capitalization.
bestätigt 10.11.2026such as one that authenticates with an API key
bestätigt 10.11.2026A session that doesn’t fetch feature flags, such as one on Amazon Bedrock
bestätigt 10.11.2026A session in bare mode or one you start with --safe-mode
bestätigt 10.11.2026
Skill im Glossar02 Die erste Sitzung10 Texte und Dateien in Ordnern11 Eine Wissensbasis aufbauen12 Hooks und Skills13 Subagenten und Schleifen17 Skills und Befehle24 Test und Spezifikation zuerst25 Prüfen und Review31 Der Harness