coding-skills

Coding Skills

<p align="center"> <b>Minimalist AI coding discipline for Claude, Kimi, Codex & CodeBuddy</b><br> One-line install · Auto-detect · Zero config </p>

<p align="center"> <a href="#installation">⚡ Quick Install</a> · <a href="#skills">📦 Skills</a> · <a href="#how-it-works">🧠 How It Works</a> </p>

---

Why

AI coding agents are capable — but undisciplined. They over-engineer, refactor unrelated code, hide confusion, and ship "done" without verifying.

Coding Skills gives your agent a set of lightweight, behavior-level rules that enforce:

• Clarity — State assumptions, surface tradeoffs, ask when uncertain
• Simplicity — No speculative features, no single-use abstractions
• Precision — Touch only what you must, match existing style
• Accountability — Define verifiable goals before writing code

Inspired by Andrej Karpathy's observations on LLM coding pitfalls and the community's enthusiasm for structured agent behavior.

---

Installation

One-liner (Recommended)

curl -fsSL https://raw.githubusercontent.com/koazy0/coding-skills/main/install.sh | bash

Or clone manually

git clone https://github.com/koazy0/coding-skills.git
cd coding-skills
bash install.sh

The installer auto-detects which AI coding tools you have and installs skills to all of them.

Add project-level baseline

cp CLAUDE.md /path/to/your-project/CLAUDE.md

This gives your agent universal coding guidelines before any skill-specific rules kick in.

---

Supported Platforms

| Platform | Global Path | Status | |----------|-------------|--------| | Claude Code | ~/.claude/skills/ | ✅ Supported | | Kimi Code CLI | ~/.kimi/skills/ | ✅ Supported | | OpenAI Codex | ~/.codex/skills/ | ✅ Supported | | CodeBuddy | ~/.codebuddy/skills/ | ✅ Supported |

> Kimi Code CLI automatically merges skills from .kimi/skills, .claude/skills, and .codex/skills — so installing to any one of them works.

---

Skills

Each skill is intentionally under 30 lines. No filler. No tutorials. Just rules that stick.

| Skill | When It Triggers | What It Enforces | |-------|-----------------|------------------| | karpathy-style | Writing, reviewing, or refactoring code | Think before coding. Simplicity first. Surgical changes only. | | code-audit | Code review, PRs, diffs | Correctness, scope control, simplicity, security checks. | | commit-style | Git commits, branches, merges | [feat] prefix, 72-char subjects, rebase-only merges. | | feature-workflow | New features, tests, merges | TDD flow: branch → red test → green → rebase → merge. | | project-init | Creating new projects | Scaffold from template, global rename, go build ready. | | api-doc-style | API endpoints, Go handlers | Swagger annotations. No hardcoded hosts. No nil arrays. |

---

How It Works

AI coding tools scan SKILL.md files at startup. When your request matches a skill's description, the full rules are loaded into context.

User: "review this PR"
      ↓
Agent: scans descriptions → matches "code-audit"
       ↓
Agent: loads code-audit/SKILL.md → follows audit rules

Because only the description (a single line) is loaded at startup, you can install all 6 skills and pay near-zero token cost until one fires.

---

Project Structure

.
├── install.sh              # Auto-detect platforms & install
├── CLAUDE.md               # Universal coding baseline
└── skills/
    ├── karpathy-style/
    ├── code-audit/
    ├── commit-style/
    ├── feature-workflow/
    ├── project-init/
    └── api-doc-style/

---

Contributing

Found a skill that doesn't trigger when it should? Open an issue.

Have a new coding discipline that fits the minimalist style? PRs welcome.

1. Fork the repo 2. Create a new skill in skills/<your-skill>/SKILL.md 3. Keep it under 30 lines 4. Submit a PR

---

Acknowledgments

• Andrej Karpathy for his public observations on LLM coding pitfalls that shaped the core principles
• The community behind andrej-karpathy-skills (MIT License) for demonstrating the impact of concise agent instructions

---

License

MIT