claude-code-settings (feiskyer)

A Claude Code settings + skills + subagents collection (1,525 stars) by Kubernetes contributor Feilong (feiskyer), designed as a multi-model gateway setup using LiteLLM Proxy pointing at GitHub Copilot. The repo ships 15 skills (including a Chinese-language deep-research orchestrator, a codex-skill for OpenAI delegation, a nanobanana image-generation skill, a Kiro-workflow skill, and a spec-kit-skill), 7 subagents, and a detailed settings.json with explicit multi-model env vars (claude-sonnet-4.6 for planning, claude-opus-4.6 for complex tasks, gpt-5-mini for Haiku-class tasks). Unlike most Claude Code frameworks, feiskyer's is fundamentally a LiteLLM → GitHub Copilot proxy configuration that makes Claude Code available without direct Anthropic API billing.

differs_from_seeds: Closest to superpowers (SKILL.md format, skills-primary architecture), but feiskyer's distinctive delta is the LiteLLM proxy multi-model gateway — routing claude-opus-4.6 for planning, claude-sonnet-4.6 for coding, and gpt-5-mini for lightweight tasks, all served through GitHub Copilot. No seed does multi-model routing through a local proxy server.

Overview — claude-code-settings (feiskyer)

Origin

Created by Feilong (feiskyer), a Kubernetes contributor. 1,525 stars, 222 forks. Last commit: 2026-04-25. Python-primary project with TypeScript skills.

Philosophy (verbatim from README)

"A curated collection of Claude Code settings, skills and sub-agents designed for enhanced development workflows. This setup includes specialized skills and subagents for feature development (spec-driven workflow), code analysis, GitHub integration, and knowledge management."

Multi-model configuration (verbatim from settings.json)

{
  "model": "opusplan",
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:4000",
    "ANTHROPIC_AUTH_TOKEN": "sk-dummy",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-4.6",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4.6",
    "CLAUDE_CODE_SUBAGENT_MODEL": "claude-sonnet-4.6",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gpt-5-mini",
    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
    "DISABLE_TELEMETRY": "1"
  }
}

LiteLLM proxy setup

The default config routes Claude Code to a local LiteLLM proxy (port 4000) that authenticates via GitHub Copilot. Users with a Copilot subscription can use Claude models without direct Anthropic billing.

Notable skills

• deep-research — Chinese-language parallel multi-agent research orchestrator using claude -p subprocesses
• codex-skill — Delegate tasks to OpenAI Codex/GPT models
• kiro-skill — Kiro spec-driven workflow
• spec-kit-skill — Spec-Kit workflow
• nanobanana-skill — Image generation (referenced but details not fetched)

Related repos

• feiskyer/codex-settings — companion repo for OpenAI Codex settings

Architecture — claude-code-settings (feiskyer)

Distribution

Claude Code Plugin marketplace + npx skills CLI.

Install

# Via Claude Code Plugin
/plugin marketplace add feiskyer/claude-code-settings
/plugin install claude-code-settings

# Individual skills via plugin
/plugin install codex-skill
/plugin install autonomous-skill
/plugin install nanobanana-skill
/plugin install kiro-skill
/plugin install spec-kit-skill
/plugin install youtube-transcribe-skill

# Via npx skills
npx -y skills add -l feiskyer/claude-code-settings     # list
npx -y skills add --all feiskyer/claude-code-settings  # install all
npx -y skills add feiskyer/claude-code-settings        # interactive

# Manual (to use as ~/.claude directly)
git clone ... ~/.claude
pip install -U 'litellm[proxy]'
litellm -c ~/.claude/guidances/litellm_config.yaml

Required runtime

• Python 3 + litellm (for LiteLLM proxy)
• GitHub Copilot subscription (for proxy auth)
• Node.js (for npx skills)
• tmux (optional, for background litellm)

Directory tree

.
├── .claude-plugin/        # Plugin marketplace config
├── agents/                # 7 subagent definitions
├── config.json            # Plugin config
├── guidances/
│   └── litellm_config.yaml  # LiteLLM proxy config
├── hooks/
│   └── hooks.json         # Stop hook for autonomous-skill
├── plugins/               # Installable plugins
│   ├── autonomous-skill/
│   ├── codex-skill/
│   ├── kiro-skill/
│   ├── nanobanana-skill/
│   ├── spec-kit-skill/
│   └── youtube-transcribe-skill/
├── scripts/               # Various scripts
├── settings.json          # Multi-model settings
├── settings/              # Additional settings
├── skills/                # 15 skill definitions
└── status-line.sh         # Status line script

Target AI tools

Claude Code (primary). The codex-skill bridges to OpenAI Codex. The kiro-skill bridges to Kiro IDE workflow.

Config files

• settings.json (multi-model config)
• hooks/hooks.json
• guidances/litellm_config.yaml

Components — claude-code-settings (feiskyer)

Skills (15 total)

Subagents (7 total)

Hooks (1 event)

Plugins (6 installable)

autonomous-skill, codex-skill, kiro-skill, nanobanana-skill, spec-kit-skill, youtube-transcribe-skill

Scripts

• status-line.sh — custom status line
• autonomous-skill/hooks/stop-hook.sh — Stop lifecycle hook

MCP servers

.mcp.json exists; configured for Exa search and Firecrawl (used by deep-research skill).

Uniqueness — claude-code-settings (feiskyer)

differs_from_seeds

Closest to superpowers (SKILL.md format, skills-primary, Claude Code plugin system) but feiskyer adds three architectural layers no seed has: (1) LiteLLM proxy multi-model routing through GitHub Copilot, (2) Chinese-language parallel research orchestration with claude -p subprocesses, and (3) cross-framework skill bridges (kiro-skill, spec-kit-skill, codex-skill). The multi-model gateway is the unique structural contribution — superpowers uses a single model, feiskyer explicitly routes claude-opus-4.6 for planning, claude-sonnet for subagents, and gpt-5-mini for lightweight tasks.

Positioning

"Skills + LiteLLM gateway for GitHub Copilot users." The primary audience appears to be developers with GitHub Copilot subscriptions who want to use Claude models through Copilot's licensing rather than direct Anthropic API billing. The deep-research skill in Chinese suggests a Chinese developer community focus.

Observable failure modes

• LiteLLM proxy setup complexity: requires pip install litellm[proxy] + GitHub Copilot auth + background process management.
• Model availability assumptions: gpt-5-mini for Haiku-class tasks assumes model availability via LiteLLM → Copilot that may change.
• Chinese-only deep-research: the deep-research skill is entirely in Chinese (description + instructions), limiting accessibility for non-Chinese readers.

Cross-references

• Explicitly references feiskyer/codex-settings (companion repo for Codex)
• Contains kiro-skill (wraps Kiro workflow) and spec-kit-skill (wraps Spec-Kit) — cross-framework bridges
• feiskyer's settings appear to be a derivative/extension of ECC (same batch) based on structural similarities