core-plugin/skills/plugin-advisor/SKILL.md
Suggests uninstalled marketplace plugins on friction. Triggered passively. Use when agent retries, user corrects approach, or no good tool exists. Out of scope: plugin installation, plugin development, MCP server setup.
npx skillsauth add bernatmv/ai-rules plugin-advisorInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
plugin-advisor skill loaded.
By Joshua Martell (jmartell). Originally from claude-kit.
A compact plugin index lives at ~/.claude_files/plugin-index.md. Each line is
one plugin with its source and a short description. The index is built by a
Sonnet subagent reading marketplace.json — no Python script needed.
On load, check freshness and rebuild if the index is missing or stale (see Freshness Check). This ensures the index is warm before friction occurs.
Use ls -l (not stat — stat triggers quoted-chars permission prompts):
ls -l ~/.claude/plugins/marketplaces/*/.claude-plugin/marketplace.json
Then read the index (~/.claude_files/plugin-index.md) and compare each marketplace
section's <!-- source: metadata against the ls -l output. Rebuild if:
ls -l but has no section in the index (new marketplace)Consider suggesting a plugin when you observe two or more of these signals:
Do not suggest on first attempt. Only after genuine friction.
~/.claude_files/plugin-index.md~/.claude/plugins/marketplaces/*/plugins/<name>/ (Anthropic first-party)
~/.claude/plugins/marketplaces/*/external_plugins/<name>/ (vendor/external)Brief, one line, with install command:
Plugin available:
playground— self-contained HTML explorers with live preview. Install:claude plugin install playground
Then continue working on the task. The suggestion is informational — don't stop.
~/.claude/plugins/installed_plugins.jsonjira skill which covers this" rather than suggestingWhen the index is stale or missing, launch a Sonnet subagent with this prompt.
Pass the ls -l output as context so the subagent can embed the metadata.
Build a compact plugin index. Use ONLY Read and Write tools. No Bash.
For EACH marketplace directory under ~/.claude/plugins/marketplaces/:
Read: ~/.claude/plugins/marketplaces/<name>/.claude-plugin/marketplace.json
Skip rules:
- Skip plugins with "-lsp" in the name
- Skip "example-plugin"
For each kept plugin, write one line:
name (source): description (under 70 chars, specific to when it helps)
Source labels:
- "Anthropic" if source starts with "./plugins/"
- "vendor" if source starts with "./external_plugins/"
- Author/org name if source is a URL
Drop any plugin whose description is too vague to match a specific problem.
Output format — one section per marketplace, with source metadata from the
ls -l output I'm providing:
# Available Plugins
# On friction, suggest matching plugin: claude plugin install <name>
## claude-plugins-official
<!-- source: 548512 Mar 13 14:23 marketplace.json -->
plugin-name (Anthropic): description
another-plugin (vendor): description
## devex-claude-plugins
<!-- source: 12345 Mar 12 09:15 marketplace.json -->
some-plugin (vendor): description
The <!-- source: SIZE DATE TIME marketplace.json --> line stores the size and
mtime from ls -l. Use the actual values from the ls -l output below:
{paste ls -l output here}
development
Keyword research and validation with real search-demand data — never ship keywords from intuition alone. Probes Google Autocomplete per language (free, no account) to prove demand and discover the exact phrasing people type, checks SERPs for winnability, and uses Keyword Planner/Ahrefs/Semrush exports when the user has access. Activates when: choosing or reviewing SEO keywords, meta keywords, page titles, article topics or slugs, landing page copy targeting search, App Store/ASO keyword fields, multilingual keyword sets, 'what should we rank for', 'keyword analysis', or auditing why a page doesn't rank. Also invoke it as a validation pass whenever another skill or task produces a keyword list.
development
Automate YooAsset hot-update and asset bundles — build bundles, run Editor simulate builds, manage Collector groups, analyze BuildReport, and validate runtime. Use when building or simulating YooAsset bundles, configuring collectors, or validating hot-update assets, even if the user just says "热更" or "打AB包". 自动化 YooAsset 热更新与资源包(构建 bundle、编辑器模拟构建、管理 Collector 分组、分析 BuildReport、运行时校验);当用户要构建或模拟 YooAsset 资源包、配置 collector、或校验热更资源时使用。
development
Source-anchored design rules for YooAsset v2.3.18 — initialization, default-package shortcuts, play modes, asset handles, loading, updates, filesystem, build, and pitfalls. Use when writing or reviewing YooAsset code, initializing packages, loading assets via handles, setting up hot-update/download, or choosing a play mode, even if the user just says "热更" or "资源包". 为 YooAsset v2.3.18 提供源码锚定的设计规则(初始化、默认包快捷方式、运行模式、资源句柄、加载、更新、文件系统、构建、陷阱);当用户要编写或审查 YooAsset 代码、初始化 package、用句柄加载资源、配置热更/下载、或选择运行模式时使用。
data-ai
Last-resort guidance for safely hand-editing Unity serialized YAML (.unity/.prefab/.asset/.meta/ProjectSettings) — reference/fileID repair, GUID safety, and merge-conflict fixes. Use when REST cannot reach the change and YAML must be hand-edited — fixing m_Script GUIDs, broken fileID references, .meta files, or merge conflicts, even if the user just says "场景文件打不开" or "引用丢了". 安全手编 Unity 序列化 YAML(.unity/.prefab/.asset/.meta/ProjectSettings)的最后手段(引用/fileID 修复、GUID 安全、合并冲突修复);当 REST 无法触达、必须手编 YAML 时使用——修复 m_Script GUID、断裂 fileID 引用、.meta 文件或合并冲突。