yao-skills
English | 繁體中文
A small, opinionated set of Claude Code skills for code review, incident triage, workflow routing, and project health checks — plus a curated subset of oh-my-claudecode skills for planning and orchestration, and behavioral guidelines from andrej-karpathy-skills.
This is a community-shared version. The methodology and decision frameworks are the core value — they transfer across projects and tech stacks. Specific tool names (Codex, Gemini CLI, etc.) are examples; adapt them to your own multi-model setup. Fork it, modify it, make it yours. If you build something useful on top of it, PRs and issues are welcome.
Skills
Author-original
| Skill | What it does |
|---|
triple-review | Orchestrator-aware three-reviewer PR review: Claude Code uses codex + agy/gemini + secondary endpoint; Codex uses claude + agy/gemini + secondary endpoint; includes severity triage, TDD fix cycle, auto-merge gate |
first-principles | Assumption-audit and incident triage discipline: 5-question audit, ground-truth verification, mandatory dual/triple review on hotfixes |
workflow-routing | Pick A/B/C/D/E/Mini workflow per task type, risk level, and current Opus / Codex quota |
project-status-review | Generate a comprehensive project status report — code stats, branch divergence, blockers, prioritized next steps |
context-hygiene | Manage session context cost: when to /compact vs handover-doc + /clear, the cache cost math (cached input is 0.1x not zero; output never cached), handover template, loop session checkpointing, and task-to-tool routing (Sonnet/Opus/Codex/Gemini CLI/secondary endpoint) |
distilled-caveman-lite-accuracy | Lite response compression — removes filler and pleasantries while preserving 100% technical accuracy. Keeps qualifiers, code identifiers, versions, step order, safety context. Safety fallback auto-expands for destructive ops, auth, crypto, compliance. Trigger: "caveman-lite", "lite mode", "brief but accurate", "less tokens" |
tc-review | Review Traditional Chinese output as a Taiwan engineer — catches Simplified Chinese character/vocabulary leaks, mid-sentence English shortcuts, jargon without gloss. Preserves original technical terms. Trigger: "繁中檢視", "tc review", "台灣用語檢查" |
Curated from oh-my-claudecode (MIT, see NOTICE.md)
ralph, plan, deep-interview, deep-dive, learner, skillify, sciomc, autoresearch, ralplan, ai-slop-cleaner, team, release, autopilot, ultrawork.
Curated from andrej-karpathy-skills (MIT, see NOTICE.md)
| Skill | What it does |
|---|
karpathy-guidelines | Behavioral guidelines distilled from Andrej Karpathy's observations on LLM coding pitfalls: think before coding, simplicity first, surgical changes, goal-driven execution |
When to use which orchestration mode
| Mode | Operating style | Best for |
|---|
autopilot | Independent autonomous single-lead agent | Fast independent feature dev / prototyping from a 2–3 line idea |
team | 5-stage pipeline (plan → prd → exec → verify → fix) | Multi-file changes needing peer architecture review |
ralph | Persistent self-referential strict-verification loop | Critical prod bug fixes that must be fully resolved |
ultrawork | Max-parallel non-team agent operation | Large-scale refactor across unrelated codebases |
ralplan | Consensus planning gate before execution | Vague / ambiguous "ralph this" / "autopilot this" requests |
Picker order: vague request → ralplan first. Independent prototype → autopilot. Multi-file design-sensitive change → team. Must-fix prod bug → ralph. Parallel-friendly bulk refactor → ultrawork.
Hooks
Hooks are shell scripts that run automatically before or after Claude Code tool calls. Unlike skills (which guide LLM behavior through prompts), hooks enforce quality gates at the infrastructure level.
| Hook | What it does |
|---|
tc-quality-hook | PreToolUse gate for AskUserQuestion — blocks Chinese text and forces the model to self-review for Simplified Chinese leakage, hallucinated characters, and mainland vocabulary before the user sees it. No external dependencies (no OpenCC, no mapping tables). |
See each hook's README for install instructions. Hooks require adding entries to ~/.claude/settings.json and restarting Claude Code.
Install
Claude Code (native — recommended)
Two paths:
A. One-shot marketplace install (easiest)
/plugin marketplace add solitude6060/Yao-skills
/plugin install yao-skills@yao-skills
Open a new Claude Code session and all 22 skills become invocable via the Skill tool / /yao-skills:<skill-name>.
B. Per-skill copy (if you only want some)
