A Curated Claude Code Kit: 12 Skills, 6 Agents, 11 Provider Presets

A fresh Claude Code install is capable but generic: no research workflow, no image generation, default model routing, and a settings file you must assemble from blog posts. feiskyer/claude-code-settings — a 2k-star, MIT-licensed repository with a Chinese-first README — packages one maintainer's whole setup into an installable plugin and a folder of configuration templates.

Why This Skill Matters

The main plugin carries 12 skills, and the README names them all: brainstorming, codex-skill, deep-research, github-fix-issue, github-review-pr, gpt-image-skill, grill-me, handoff, nanobanana-skill, skill-creator, translate, and youtube-transcribe-skill. Around them sit 6 specialized subagents (pr-reviewer, github-issue-fixer, instruction-reflector, deep-reflector, insight-documenter, ui-engineer) and an agents/ directory that extends Claude Code where a skill alone would not.

The configuration half is the quieter draw: the settings table lists 11 provider presets — GitHub Copilot, LiteLLM, DeepSeek v3.1, Qwen3-Coder-Plus, Kimi-K2-Instruct on SiliconFlow, Vertex AI, Azure AI, Azure AI Foundry, MiniMax (global and China endpoints), and OpenRouter. Each is a settings.json pre-wired to that provider's endpoint, so switching model backends is a file copy instead of a docs dig.

Installation

Two equivalent marketplace routes, straight from the README:

/plugin marketplace add feiskyer/claude-code-settings
/plugin install claude-code-settings

Or from the terminal:

claude plugin marketplace add feiskyer/claude-code-settings
claude plugin install claude-code-settings@claude-code-settings

One boundary to know before you start: ~/.claude/settings.json is not part of the plugin — the README marks it as manual setup. The repo's root settings.json is the author's template, defaulting to the copilot-gateway proxy at http://localhost:4141; point ANTHROPIC_BASE_URL elsewhere (LiteLLM's http://localhost:4000, say) if you use another gateway. The template's model names — claude-sonnet-5, claude-opus-5, claude-haiku-4-5 — must exist in your account, so replace them with what your gateway actually serves. npx skills add feiskyer/claude-code-settings can install individual skills, but the README warns that route may lag the repo's latest versions.

Real Workflow: Run a Deep Research Pass

deep-research is the kit's flagship. Give it a research goal — the trigger words include deep research and multi-agent parallel research — and it splits the goal into parallel subtasks, executes each through claude -p subprocesses, aggregates the results, and delivers a refined report. The strategy is skills-first: use installed skills, then MCP tools, then WebFetch/WebSearch. Runs scale from micro (1–2 tasks) to large (15+).

Everything lands in a structured output directory:

.research/<name>/
├── prompts/           # 子任务 prompt
├── child_outputs/     # 子进程输出
├── logs/              # 执行日志
├── raw/               # 原始数据缓存
└── final_report.md    # 最终报告

Because each chapter is refined and source-checked before the report is assembled, the final file reads like a brief an analyst wrote, not a chat log. When you need a second opinion on plans instead of research, grill-me runs a high-intensity interview — one question at a time, each with a recommended answer, cross-checked against your code and a maintained CONTEXT.md glossary.

Tips

• Don't install everything — the README's own FAQ: too many skills increase context consumption and reduce trigger accuracy. Install the plugin, then disable what you don't use.
• The repo's settings.json template defaults defaultMode to acceptEdits (edits auto-accepted, commands still confirmed). bypassPermissions exists but the README asks you to understand the security implications first.
• Considering the manual route? It clones the repo over ~/.claude — back up your existing directory first (mv ~/.claude ~/.claude.bak), the README says this in bold.
• GitHub Copilot as a backend needs its gateway running: npx copilot-gateway@latest start --proxy-env, then authenticate via the device-login prompt.
• Two image skills coexist by routing rule: nanobanana-skill (Gemini) is the default, gpt-image-skill fires only when you explicitly name OpenAI/GPT.

When Not to Use This

This kit is opinionated toward its author's stack — a Copilot/LiteLLM-style gateway setup, Python-based image skills with their own API keys, and workflow skills that assume Claude Code. If you run a different harness, the skills are the only portable part. And if you already maintain a mature personal ~/.claude, adopting a 12-skill bundle wholesale is how the context cost the FAQ warns about happens; cherry-pick via npx skills instead, with the version-lag caveat in mind.

See the leaderboard for more skills.