DEV Community

William Wang
William Wang

Posted on

Agent Skills in 2026: How to Customize Claude Code, Codex CLI & Gemini CLI

#ai #programming #productivity #tutorial

AI coding agents like Claude Code, OpenAI Codex CLI, and Google Gemini CLI are powerful out of the box. But they become dramatically better when you teach them your workflows through Agent Skills — simple Markdown files that encode reusable instructions.

I've been using skills daily for the past few months, and the difference is night and day. Here's everything I've learned about customizing these three major CLI tools.

What Are Agent Skills?

An Agent Skill is a Markdown file (typically SKILL.md, AGENTS.md, or GEMINI.md) that tells an AI coding agent how to perform a specific task. Think of it as a reusable prompt with structure.

Unlike MCP servers (which require running code) or simple rules files (which are just text constraints), skills sit in a sweet spot: powerful enough for complex workflows, simple enough that anyone can create one.

---
name: Code Review
description: "Thorough adversarial code review"
---

## Steps
1. Read all changed files
2. Check for bugs, security issues, performance problems
3. Output findings grouped by severity
Enter fullscreen mode Exit fullscreen mode

That's it. Drop this file in your project, and your AI agent knows how to do a thorough code review every time.

For a deeper comparison of when to use skills vs MCP vs rules, check out this complete breakdown.

Claude Code: The SKILL.md Ecosystem

Claude Code looks for SKILL.md files in .claude/commands/ or your project root. As of April 2026, the ecosystem has exploded — there are hundreds of community-built skills covering everything from deployment automation to video production.

Getting Started 

# Install a skill from a registry
claude install @tokrepo/code-review# Or just create your own
mkdir -p .claude/commands
touch .claude/commands/SKILL.md
Enter fullscreen mode Exit fullscreen mode

Best Skills I've Tested

After testing dozens of skills, here are the ones that actually made a difference in my daily workflow:

  • Adversarial Code Review — catches bugs that standard review misses
  • Token-Saving Compact — reduces context usage by 40-50%
  • Git Commit Conventions — consistent, meaningful commit messages
  • Test Generator — writes tests that actually cover edge cases
  • Documentation Writer — generates docs that stay in sync with code

I wrote a full ranked list with install commands: 15 Best Claude Code Skills in 2026.

OpenAI Codex CLI: AGENTS.md

Codex CLI uses AGENTS.md files — similar concept, slightly different format. The good news: most SKILL.md files work with Codex with minimal changes.

# Install Codex CLI
npm install -g @openai/codex# Create an agent skill
touch AGENTS.md
Enter fullscreen mode Exit fullscreen mode

Codex's strength is its sandbox execution model — it runs code in an isolated environment by default, which makes it great for skills that involve running tests or builds.

If you're getting started with Codex skills, I put together a setup guide: Getting Started with Codex CLI Skills.

Google Gemini CLI: GEMINI.md

Google's entry into the CLI agent space (launched 2026) uses GEMINI.md files. It's the newest of the three, but it already has a growing extension ecosystem with 40+ official Google-built extensions.

# Install Gemini CLI
npm install -g @anthropic-ai/gemini-cli# Create a skill
touch GEMINI.md
Enter fullscreen mode Exit fullscreen mode

Gemini CLI's advantage is tight integration with Google Cloud services — if your stack is on GCP, the built-in extensions for BigQuery, Cloud Run, and Firebase are excellent.

Full setup guide here: Gemini CLI Agent Skills Guide.

Cross-Platform Skills

Here's the interesting part: the skill formats are converging. A well-written SKILL.md works on Claude Code, Codex CLI, and Gemini CLI with zero or minimal changes. This means you can write once and use across all three tools.

The key fields that work everywhere:

  • name and description in frontmatter
  • Markdown-structured instructions
  • Step-by-step workflows
  • Code blocks with examples

If you want to learn how to create your own cross-platform skill from scratch, here's a step-by-step tutorial: How to Create Your First Agent Skill.

Where to Find Skills

Building your own skills is great, but there's no need to start from scratch for common workflows. Several registries have emerged:

  • TokRepo — 600+ curated skills, MCP configs, and prompts with one-click install
  • GitHub — search for SKILL.md or AGENTS.md files
  • Claude Code's /powerup — built-in tutorial that teaches skill basics

The Token Cost Problem (and How Skills Solve It)

One thing nobody talks about enough: AI coding agents are expensive. Claude Opus costs $15/$75 per million tokens. A heavy coding session can easily burn through $20-50.

Skills help because they:

  1. Replace repetitive prompts — encode once, reuse forever
  2. Reduce context size — focused instructions instead of rambling conversations
  3. Minimize iteration — better first-pass results mean fewer back-and-forth cycles

I measured the difference: How to Cut Claude Code Token Costs by 50%.

What's Next

The agent skills ecosystem is still early but growing fast. The convergence across Claude Code, Codex, and Gemini CLI means that skills you write today will likely work with whatever new tool comes out tomorrow.

If you're not using skills yet, start with one. Install a code review skill, use it for a week, and see the difference. Then create your own for your most repetitive workflow.

All the skills and guides mentioned in this article are available on TokRepo — a registry of 600+ curated AI assets for coding agents.

Top comments (0)