Top 10 Skills for Claude Code and Codex CLI

The real skill isn’t getting AI to answers! But to do so in a manner that fits our budgets and fulfils our requirements. It’s guiding it with clear context and turning its output into useful action.

This list is built around a simpler idea. Instead of searching through thousands of skills, you start with the exact problem you’re facing. Each common frustration is matched to a specific skill, so you can install only what you need and make tools like Claude Code or Codex CLI more useful.

1. For AI/software development

obra/superpowers | ★ 267,733 · 23,919 forks

Moreover, an agentic skills framework & software development methodology that works. Instead of jumping into code, the agent asks you questions until it understands what you want, shows you the plan in pieces short enough to read, and waits for your approval before building anything.

What makes this skill special?

• It asks first and builds second, which is the opposite of the default behaviour
• It writes tests as it goes, rather than promising them later
• Once approved, it can work on its own for hours without drifting off the plan
• Nothing to remember after installing. It kicks in by itself.

Works with: Both Claude Code and Codex CLI.

In fact, perfect for anyone who has watched an agent confidently build the wrong thing.

2. For agent harnesses

affaan-m/ECC | ★ 238,146 · 36,166 forks

If everyone on your team has their agent configured differently and nobody can reproduce anyone else’s results, one skill won’t help. You need a shared setup. Consequently, this is that: a large, opinionated bundle covering planning, code review, build fixes, security, and architecture, all installed in one go.

What makes this skill special?

• It’s built around roles (the planner, the reviewer etc.) rather than a pile of unrelated tools
• In addition, a built-in security scanner checks your own setup for leaked secrets and risky settings
• You choose which coding standards apply to which projects, instead of forcing them everywhere
• The documentation is honest about which features work on which tools

Works with: Best with Claude Code. Meanwhile, Codex is supported through a separate setup step.

Best for teams, not individuals.

3. For getting AI to ask clarifying questions

multica-ai/andrej-karpathy-skills | ★ 200,098 · 20,585 forks

Your agent quietly decides what you meant, runs with it, and you find out twenty minutes later. For example, this repo has 200,000 stars for a single file that targets exactly that. Similarly, it takes Andrej Karpathy’s widely shared post about where coding agents go wrong and turns each complaint into a rule.

What makes this skill special?

• It’s one file with nothing else attached, so it’s the cheapest thing on this list to try
• The fourth rule is the one that changes results. “Fix the bug” becomes “write a test that shows the bug, then make it pass.” Give the agent something it can check itself against and it stops guessing.
• It tells you how to know it’s working with questions arriving before the work rather than apologies after!

Works with: Claude Code and Cursor out of the box.

The first thing most people should install.

4. For getting shorter, simpler code

DietrichGebert/ponytail | ★ 97,139 · 5,337 forks

You ask for a date picker. While, your agent installs an extra library, writes a wrapper for it, adds a stylesheet, and starts a discussion about timezones. Instead, this skill gives you the single line of built-in code that already does the job. Nevertheless, the idea is the senior engineer who looks at your fifty lines, says nothing, and replaces them with one.

It’s also the only entry here that published a test of itself against itself. In short, twelve real feature requests on a real project, with and without the skill:

What makes this skill special?

• As a result, roughly half the code, for a fifth less money and a quarter less waiting
• In fact, simply telling the agent “keep it short” gets you part of the way and quietly drops a safety check doing it. This doesn’t.
• Consequently, the numbers are honest about their limits: the saving is huge where the agent was about to over-build, and near zero where the code was already simple.

Works with: Claude Code and Codex CLI.

Perfect if you’re the one who has to maintain the result.

5. For saving tokens

JuliusBrussee/caveman | ★ 96,300 · 5,500 forks

Agents pad with sentences like “Sure! Furthermore, I’d be happy to help with that. The issue you’re experiencing is most likely…” you pay for every word of that! In contrast, this skill makes the agent talk like a caveman. Same answer, a third of the words.

Code, commands, paths and error messages come through untouched. Only the words gets squeezed.

What makes this skill special?

• Keeps your language. Write Portuguese, get caveman response in Portuguese.
• Also shortens commit messages and code review comments
• Honest about limits: savings shrink on unattended coding work, sometimes to nothing
• JetBrains tested it across 86 tasks and found answer quality unchanged

Works with: Claude Code and Codex CLI supported.

Perfect if you’re watching your usage limits.

Caveman or ponytail?

Both promise savings but at opposite ends. Ponytail shrinks the code. Caveman shrinks the words around it.

6. For recalling previous sessions

thedotmack/claude-mem | ★ 89,815 · 7,819 forks

If you start every session re-explaining last week, this fixes it. It quietly records what happens as you work, stores it on your own machine, and lets the agent look things up later by meaning rather than exact wording. Ask clarifying questions like “what did we decide about the login flow?” and it finds the answer.

What makes this skill special?

• Most memory tools fail by dumping everything back into the conversation and slowing it to a crawl. This one searches first and fetches only what’s relevant.
• Everything stays local, on your machine
• It comes with a simple web page where you can read your own history back
• The agent decides when to look something up, so you don’t have to prompt it

Works with: Claude Code only.

Best for projects that run for weeks, not one-off tasks.

7. For Agent workflows

addyosmani/agent-skills | ★ 82,313 · 8,844 forks

If you’d like a professional workflow but don’t want to design one yourself, this hands you a ready-made one. It breaks work into six stages you trigger by name: define what you’re building, plan it, build it, test it, review it, ship it. Two extras cover speeding up a slow website and simplifying messy code.

What makes this skill special?

• Each stage carries a principle rather than a checklist: agree what you’re building before you build it, small pieces one at a time, tests as proof
• It picks the right stage by itself based on what you’re doing, so you can also just work normally
• There’s an automatic mode that runs the whole build in one pass after you approve the plan. It still tests everything and still stops if something breaks.

Works with: Claude Code and Codex CLI, plus 70+ other tools.

Perfect for solo developers who want structure without inventing it.

8. For mapping unfamiliar codebase

Egonex-AI/Understand-Anything | ★ 77,730 · 6,526 forks

You’ve just joined a team and the project is 200,000 lines long. Every other entry here improves how the agent writes. This one improves how you read. Point it at a project and it produces an interactive map of how everything connects, which you explore in your browser.

What makes this skill special?

• Click any part of the map and get a plain-English explanation of what it does and what depends on it
• The stated goal is a map that teaches you how the pieces fit, not one that shows off how complicated the project is
• There’s a live demo on their site, so you can see what you’re getting before installing anything
• Just as useful before a big cleanup, when what you need is a map rather than a list of changes

Works with: Claude Code, Codex CLI, and about ten other tools.

Best for inherited projects and getting oriented.

9. For retaining context in conversation resets

OthmanAdi/planning-with-files | ★ 26,001 · 2,178 forks

Every agent forgets. Clear the chat, hit a limit, or close your laptop, and the plan goes with it. This skill writes the plan, the findings, and the progress into three ordinary files on your computer, then feeds them back in every time you type. The conversation resets. The plan doesn’t.

What makes this skill special?

• It happens automatically, so nobody has to remember to save anything
• An optional final check stops the agent declaring the job done before it actually is
• Three plain text files you can open and read yourself, not a hidden database
• Their own testing shows an agent picking up where it left off in about 5 exchanges instead of 13. Worth knowing those tests were run by the author, not independently.

Works with: Claude Code, Codex CLI, and 60+ other tools.

Perfect for work measured in days, and the best option for Codex users who want memory.

10. For checking if a skill is safe

NVIDIA/SkillSpector | ★ 14,278 · 1,194 forks

Install this before anything else on this list, including this list. A skill is a set of instructions your agent will follow, with access to your files and your accounts. Research NVIDIA cites found that “roughly one in four skills contains a security flaw, and one in twenty looks deliberately malicious”.

This is a free scanner from NVIDIA that reads a skill before you install it and tells you whether it’s safe.

What makes this skill special?

• Checks against 68 known bad patterns, including skills that quietly steal data or hand themselves more access than they should have
• Give it a link, a folder, or a single file. It handles all three.
• You get a simple score out of 100 and a plain summary, not a wall of security jargon
• Run it again later and it only flags what’s new

Works with: Everything. It’s a scanner, not a skill, so it doesn’t care which agent you use.

For everyone, and especially anyone about to install a skill from someone they’ve never heard of.

How to install Claude/Codex Skills

Most of these take under a minute. Two commands cover almost everything.

In Claude Code, type this in the chat:

/plugin marketplace add <owner>/<repo>
/plugin install <name>

In Codex CLI, run this in your terminal:

codex plugin marketplace add <owner>/<repo>
codex plugin add <name>

There’s also a universal installer that works across 70+ tools if you’d rather not remember which is which:

npx skills add <owner>/<repo>

If you use both Claude Code and Codex, runkids/skillshare keeps one set of skills in sync across both, so you’re not maintaining two copies.

Where to start?

There’s no single skill that can be the best for all (though I’d recommend the Nvidia profile inspector to all.

If your agent keeps writing far more code than the job needed, start with andrej-karpathy-skills or ponytail. If the problem is that there’s no process at all, agent-skills hands you one, and superpowers goes further and rebuilds how you work from the ground up.

Whatever you choose, scan it with SkillSpector first. Three or four skills is the practical limit anyway. Past that they start overlapping and you lose track of which one is doing what.

Read more: Top 5 Agentic Skill for Writing

Frequently Asked Questions