Top 10 Claude Code Skills to Operate and Extend

Best for: Auditing CLAUDE.md memory files for completeness, currency, and actionable workflow instructions

claude-md-improver audits CLAUDE.md, .claude.md, and local memory files across project roots and monorepo packages. It evaluates build commands, architecture notes, non-obvious conventions, and instruction actionability against a structured rubric.

This skill instructs the agent to present an evaluation report with recommended additions and diffs before seeking user approval. While prompt instructions ask for approval before editing, this is behavioral guidance for the LLM rather than a hard mechanical enforcement barrier in the runtime.

• Monorepo and root CLAUDE.md discovery
• Structured rubric scoring commands and architecture clarity
• Generates quality evaluation report with recommended changes
• Instructs agent to ask for user approval prior to editing

Best for: Setting up or auditing Claude Code settings.json permission allowlists tailored to repository stacks

claude-settings-audit scans package manifests, lockfiles, and configuration to detect active tech stacks (Node, Python, Go, Rust, Ruby, Java, monorepo tools) and produce recommended settings.json permission rules.

Be mindful that the generated baseline includes broad Bash wildcard allowlists such as Bash(cat:*), Bash(find:*), or Bash(gh api:*). Because prefix wildcards match commands regardless of destructive arguments or output redirects, they do not guarantee read-only execution at the OS or API level. Review and narrow the suggested allowlists before saving them to .claude/settings.json.

• Automatic stack and dependency manifest detection
• Generates project-tailored settings.json allowlists
• Wildcard prefix rules bypass prompts without strict OS read-only limits
• Requires careful manual review before adding to configuration

Best for: Preventing Claude Code from accidentally running git push, hard resets, forced clean, or branch deletions

git-guardrails-claude-code sets up PreToolUse protection to block hazardous git actions like git push (including --force), git reset --hard, git clean -f, git branch -D, and indiscriminate checkout or restore commands.

Protection is strictly opt-in and does not take effect merely by installing the skill. You must explicitly copy block-dangerous-git.sh to .claude/hooks/ (or ~/.claude/hooks/), grant execute permissions via chmod +x, and register the hook in .claude/settings.json.

• Blocks git push, reset --hard, clean -f, and branch -D
• Requires copying block-dangerous-git.sh and making it executable
• Opt-in configuration via PreToolUse in .claude/settings.json
• Supports project-level or global home-directory scoping

Best for: Investigating token consumption, cache hit ratios, subagent overhead, and expensive prompt spikes

session-report aggregates Claude Code usage by running its bundled analyze-sessions.mjs script against transcripts in ~/.claude/projects, extracting overall token totals, by_project, by_subagent_type, by_skill, cache_breaks, and top_prompts into template.html.

Because top_prompts includes raw prompt strings and project folder names from your local sessions, exercise discretion before sharing the generated HTML report outside your immediate team.

• Analyzes local transcripts into overall, by_project, by_skill, and cache_breaks data
• Embeds metrics and top_prompts into an interactive standalone template.html
• Tracks token spikes, subagent calls, and cache-break patterns
• Report includes prompt text, project paths, and token counts from local sessions

Best for: Identifying token bloat and reclaiming context window headroom when sessions degrade or slow down

context-budget inventories loaded project assets to locate context overhead. It calculates token estimates across agent definitions, skill directories, rule files, MCP tool schemas, and CLAUDE.md memory files.

This skill uses a heuristic estimation formula (word count × 1.3 for markdown files, ~500 tokens per tool schema for MCP servers) rather than an exact BPE tokenizer. In addition, standalone installation via npx skills add does not create a native /context-budget slash command in Claude Code; it runs as standard skill instructions unless explicitly mapped to a custom command.

• Estimates context window overhead across project components
• Heuristic sizing: words × 1.3 for text, ~500 tokens per MCP tool
• Flags heavy agent files, bloated frontmatter, and duplicate rules
• Does not register a standalone /context-budget slash command

Best for: Discovering relevant Claude Code automations without manually researching hundreds of plugin options

claude-automation-recommender inspects project dependencies, test harnesses, database configurations, and issue trackers to recommend targeted Claude Code automations across hooks, subagents, skills, and MCP servers.

The skill instructions are designed to be advisory—analyzing tools and suggesting 1–2 high-leverage options per category rather than modifying files. However, non-modification is a behavioral design intent of the prompt instructions, not a runtime sandbox guarantee.

• Detects frontend, backend, database, and testing tools
• Recommends focused hooks, MCP servers, and subagents
• Designed as an advisory scan that does not attempt file edits
• Prioritizes 1–2 high-value suggestions per category

Best for: Scaffolding, refining, and validating custom SKILL.md workflows with rigorous testing and evaluation loops

skill-creator guides developers through the complete authoring lifecycle for agent skills: capturing user intent, structuring frontmatter, writing concise imperative instructions, and separating bundled scripts and references.

Running the skill's quantitative evaluation and benchmarking workflows (such as eval-viewer/generate_review.py) requires a configured Python evaluation environment with necessary dependencies. Ensure the local execution environment supports Python benchmarking before running eval loops.

• End-to-end SKILL.md authoring and frontmatter design
• Three-level progressive disclosure architecture
• Quantitative eval benchmarking and description optimization
• Requires a Python runtime environment for eval-viewer scripts

Registry detail does not report adoption for this official advanced card.

Best for: Structuring Claude Code plugins with valid .claude-plugin/plugin.json manifests and component directories

plugin-structure specifies the directory conventions, component layouts, and manifest schema for Claude Code plugins. It details how .claude-plugin/plugin.json configures commands, agents, skills, and hooks using ${CLAUDE_PLUGIN_ROOT}.

This skill is reference documentation and design guidance only—it is not an executable plugin runtime or background daemon. Use it to scaffold clean, distributable plugins that integrate predictably into Claude Code's auto-discovery mechanism.

• Standardized directory layout for commands, agents, and hooks
• Manifest schema rules for .claude-plugin/plugin.json
• Guidance-only architectural reference, not a runtime service
• Portable path references via ${CLAUDE_PLUGIN_ROOT}

Registry detail does not report adoption for this official advanced card.

Best for: Implementing user-configurable settings and state files for custom Claude Code plugins

plugin-settings documents the .claude/plugin-name.local.md convention for storing project-scoped plugin configuration. It pairs structured YAML frontmatter for boolean or string parameters with markdown bodies for prompt templates and notes.

Writing a .local.md settings file has zero effect unless a consumer plugin is actively installed and programmed to read and apply those settings. Without an active hook, command, or subagent checking the file, the settings remain inert text.

• Structured YAML frontmatter paired with markdown context
• Project-scoped .claude/plugin-name.local.md pattern
• Requires an active consumer plugin to read and apply configuration
• Bash hook and command parsing patterns for plugin authors

Registry detail does not report adoption for this official advanced card.

Best for: Writing project safety policies that intercept dangerous terminal commands or sensitive file edits

writing-hookify-rules teaches the syntax and condition structures for Hookify rules stored in .claude/hookify.{rule-name}.local.md. It explains how to match regex patterns against bash commands, file modifications, stop events, or user prompts to issue warnings or block execution.

Hookify rules evaluate pattern matches when Hookify is installed and enabled in the Claude Code runtime. While Hookify supports bash command matching (which can include git commands), this guide scopes Hookify to non-git safety policies (such as blocking rm -rf, warning on .env secrets, or checking test completion), leaving git command protection to dedicated pre-tool git hooks.

• Author .claude/hookify.{rule-name}.local.md rule files
• Multi-condition matching across bash commands, file edits, and stop events
• Requires an installed and enabled Hookify runtime to trigger
• Guide scopes Hookify to non-git safety checks alongside dedicated git hooks