Skill Finder

Search and install Claude Code skills from multiple sources.

Workflow

Follow these steps in order when the user invokes this skill.

Step 1: Detect Language and Parse Keywords

Take the search keywords from $ARGUMENTS.

• If no keywords provided, use AskUserQuestion to ask the user what kind of skill they're looking for.
• Detect the primary language of the user's keywords:
  • If keywords contain CJK characters (Chinese/Japanese/Korean), primary language is CJK
  • If keywords contain non-ASCII characters (e.g., Spanish, French, German), detect that language
  • Otherwise, primary language is English
• Store the detected language for later use in result presentation (Step 3)
• If keywords are not in English, translate them to English while keeping the originals. Generate 2-4 English keyword variants for broader matching.
  • Example: "数据库管理" → detected: Chinese → keywords: ["database", "db", "management", "sql"]
  • Example: "文件处理" → detected: Chinese → keywords: ["file", "processing", "document", "handler"]
  • Example: "Slack 通知" → detected: Chinese → keywords: ["slack", "notification", "messaging"]
• If keywords are in English, still consider adding 1-2 synonyms/related terms.
• Remove common stop words (automation, tool, plugin, skill) from matching but keep them for display.

Step 2: Search Multiple Sources

Run the following searches in parallel to maximize speed.

Source A: skills.sh Leaderboard (via skills CLI)

Run for each primary keyword (up to 3):

npx skills find <keyword> 2>&1 | head -60

Parse the output to extract:

• Skill identifier: owner/repo@skill-name
• Install count
• URL

Source B: GitHub Skill Registries (via search script)

Locate and run the TypeScript search script included with this skill (runs via npx -y bun):

SKILL_DIR=$(find ~/.claude/skills .claude/skills -name "search_github.ts" -path "*/inm-skill-finder/*" -exec dirname {} \; 2>/dev/null | head -1)
SKILL_DIR=$(dirname "$SKILL_DIR")  # go up from scripts/ to skill root
npx -y bun run "$SKILL_DIR/scripts/search_github.ts" keyword1 keyword2 --descriptions

This searches all enabled registries. Registries are configured in:

• config.yaml (recommended) — unified YAML configuration with simplified registry format
• registries.json + registries.local.json (legacy) — still supported for backward compatibility

Preferred: Edit config.yaml directly

Users can simply add GitHub repo URLs to config.yaml:

registries:
  - url: https://github.com/your-org/your-skills
    enabled: true
  - url: owner/repo  # Short format also works
    enabled: true

Copy config.example.yaml to config.yaml to get started.

Alternative: CLI commands (creates registries.local.json):

# List all registries and their status
npx -y bun run "$SKILL_DIR/scripts/search_github.ts" --list-registries

# Add a custom registry
npx -y bun run "$SKILL_DIR/scripts/search_github.ts" --add-registry owner/repo --name "My Skills"

# Remove a custom registry
npx -y bun run "$SKILL_DIR/scripts/search_github.ts" --remove-registry owner/repo

# Disable/enable a registry
npx -y bun run "$SKILL_DIR/scripts/search_github.ts" --disable-registry composio
npx -y bun run "$SKILL_DIR/scripts/search_github.ts" --enable-registry composio

If the script is not found, fall back to the skills find CLI command from Source A.

Source C: General Web Search (supplementary)

Use WebSearch to find additional skills:

[keyword] claude code skill site:github.com

Extract any GitHub repos that contain SKILL.md files and are relevant.

Step 3: Deduplicate and Present Results

1. Merge results from all sources, removing duplicates (same skill name from same repo).
2. Sort by relevance: exact matches first, then by install count (if available), then by fuzzy score.
3. Localize the presentation based on the detected language from Step 1:
   • If user used Chinese keywords: Present UI text in Chinese (table headers, prompts, instructions)
   • If user used other non-English keywords: Present in that language if you can, otherwise English
   • If user used English keywords: Present in English (default)
   • Skill names and descriptions remain in their original language (typically English from GitHub)
4. Present results in a clear numbered table format:

Example (Chinese input):

## 搜索结果："数据库管理" → [database, db, management, sql]

| #  | 技能名称                    | 来源              | 安装数   | 描述                 |
|----|----------------------------|-------------------|----------|----------------------|
| 1  | postgres-automation        | skills.sh         | 5.2K     | Automate Postgres... |
| 2  | mysql-manager              | awesome-claude... | -        | MySQL management...  |

Example (English input):

## Search Results for: database management → [database, db, management, sql]

| #  | Skill Name                  | Source            | Installs | Description          |
|----|----------------------------|-------------------|----------|----------------------|
| 1  | postgres-automation         | skills.sh         | 5.2K     | Automate Postgres... |
| 2  | mysql-manager               | awesome-claude... | -        | MySQL management...  |

5. Use AskUserQuestion with multiSelect: true to let the user choose which skills to install.
   • List top 4 options (or fewer if less results)
   • Each option: "[name] (from [source])"
   • Question text should be localized based on detected language

Step 4: Check User Preferences and Ask Installation Options

FIRST: Check if user has predefined preferences in config.yaml:

SKILL_DIR=$(find ~/.claude/skills .claude/skills -name "search_github.ts" -path "*/inm-skill-finder/*" -exec dirname {} \; 2>/dev/null | head -1)
SKILL_DIR=$(dirname "$SKILL_DIR")
PREFS=$(npx -y bun run "$SKILL_DIR/scripts/search_github.ts" --show-preferences 2>/dev/null)

Parse the JSON output to extract install_method and install_location:

• If install_method is not 'ask': Use that method directly, skip Question 1
• If install_location is not 'ask': Use that location directly, skip Question 2
• If both are 'ask' or undefined: Ask both questions

If questions are needed, use AskUserQuestion and localize based on detected language from Step 1:

Question 1 - Installation Method (ask only if preference is 'ask'):

• npx skills add (Recommended) — Uses the skills.sh CLI tool. Easiest and most standard method. Works for skills.sh listings and GitHub repos.
• Direct download — Downloads SKILL.md (and supporting files) directly via curl. Lightweight, no extra tools needed.
• git clone — Clones the full repository, then copies the skill directory. Best for skills with many supporting files.

Question 2 - Installation Level (ask only if preference is 'ask'):

• Project-level (Recommended) — Installs to .claude/skills/ in the current project. Only available in this project.
• User-level (global) — Installs to ~/.claude/skills/. Available across all your projects.

Example localized questions (Chinese):

• Question 1: "选择安装方式" with options like "npx skills add（推荐）", "直接下载", "git clone"
• Question 2: "选择安装级别" with options like "项目级别（推荐）", "用户级别（全局）"

Step 5: Execute Installation

Based on user choices, execute the appropriate installation command.

Method: npx skills add

# For skills.sh listings (owner/repo@skill format):
npx skills add owner/repo --skill skill-name          # project-level
npx skills add owner/repo --skill skill-name -g        # user-level (global)

# For standalone repos:
npx skills add owner/repo                              # project-level
npx skills add owner/repo -g                            # user-level (global)

For skills from ComposioHQ/awesome-claude-skills:

npx skills add ComposioHQ/awesome-claude-skills --skill skill-name
npx skills add ComposioHQ/awesome-claude-skills --skill skill-name -g

Method: Direct Download

# Determine target directory
TARGET="$HOME/.claude/skills/SKILL_NAME"    # user-level
TARGET=".claude/skills/SKILL_NAME"          # project-level

mkdir -p "$TARGET"

# Download SKILL.md
curl -sL "RAW_SKILL_URL" -o "$TARGET/SKILL.md"

# If there are additional files (scripts/, resources/), download those too
# Check the repo for extra files first

Method: git clone

# Determine target directory
TARGET="$HOME/.claude/skills/SKILL_NAME"    # user-level
TARGET=".claude/skills/SKILL_NAME"          # project-level

TEMP=$(mktemp -d)
git clone --depth 1 --filter=blob:none --sparse \
  "https://github.com/OWNER/REPO.git" "$TEMP/repo"
cd "$TEMP/repo"
git sparse-checkout set "SKILL_PATH"

mkdir -p "$TARGET"
cp -r "SKILL_PATH/." "$TARGET/"
rm -rf "$TEMP"

Step 6: Verify and Report

After installation:

1. Verify the SKILL.md file exists at the target location.
2. Read the installed SKILL.md and display:
   • Skill name
   • Description
   • How to invoke it (e.g., /skill-name or automatic invocation)
3. If user-invocable is not false in the skill's frontmatter, inform the user they can invoke it with /skill-name.
4. If the skill has requires in frontmatter (e.g., MCP servers), warn the user about prerequisites.
5. Localize the completion message based on the detected language from Step 1:
   • Chinese user → report in Chinese
   • Other language user → report in that language or English
   • English user → report in English

Error Handling

• No results found: Suggest the user try different keywords, broader terms, or browse https://skills.sh directly.
• GitHub API rate limit: Advise setting GITHUB_TOKEN environment variable. Fall back to skills find CLI.
• Installation failure: Show the error, suggest alternative installation method.
• Network issues: Suggest checking internet connection and trying again.

Notes

• The skills CLI (npx skills) is the primary tool from https://skills.sh and supports search, install, update, and removal.
• The search script uses a configurable registry system (registries.json + registries.local.json) with 7 default repos including ComposioHQ, Anthropic, VoltAgent, BehiSecc, Antigravity, and more.
• READMEs are fetched via raw.githubusercontent.com (no API quota consumed). If a README can't be parsed, the script falls back to the GitHub Trees API.
• Search matches against both skill names and descriptions, so searching "email" will find skills like sendgrid-automation whose description mentions email.
• Skills installed at project-level (.claude/skills/) are auto-discovered by Claude Code with live reload.
• Skills installed at user-level (~/.claude/skills/) work across all projects.