Ask for one small fix and your coding agent hands back a refactor nobody wanted. Agent skills exist to stop that. They are SKILL.md files that most coding agents load. One pack follows you from Claude Code to Codex, Cursor or Gemini CLI.
We scouted the internet for the agent skills real humans actually rate. We read Hacker News threads, YouTube reviews, blogs and GitHub issues. Every pick installs in more than one agent. Each one fixes a single failure, from building the wrong thing to pages that look AI-made.
The Rundown
- Your Agent Builds The Wrong Thing: Grill Me. Agree the plan before the build.
- Your Agent Jumps Straight Into Code: Superpowers. Write a spec and plan first for multi-file work.
- Your Agent Overbuilds Everything: Karpathy Guidelines. Make the agent think first and keep changes small.
- Your Agent Skips Tests And Review: Addy Osmani’s Agent Skills. Run the spec, test and review steps before shipping.
- Your Page Works But Looks AI-Made: Impeccable. Find and fix flaws in pages you have.
- Every New Page Gets The Same Safe Look: Frontend Design. Plan a distinct look before building a new page.
The Best AI Agent Skills, Ranked
Grill Me
Price | Free, MIT license |
Install | Plugin per agent, or npx skills add mattpocock/skills |
Platforms | Claude Code, Codex, GitHub Copilot, Gemini CLI, Cursor and OpenCode via npx skills |
Best for | Agreeing the plan before anything gets built |
The problem. You describe a small feature and the agent builds a version of it you never agreed to. The design choices get made in code, before you see them.
What it does. Grill Me makes the agent interview you before it builds. You type /grill-me followed by your plan, and a second skill called grilling takes over with numbered questions. Each question comes with a recommended answer, so yes accepts it. It looks up facts itself, leaves choices to you, and will not act until you agree on the plan. In Claude Code, keep /grill-me at the start of the prompt. Install grilling with it, because grill-me alone fails without it (GitHub issue 1319). The README has installs for Claude Code, Codex, GitHub Copilot, Gemini CLI, Cursor and OpenCode.
What people found
- Theo ran the grilling prompt on their project Lakebed. The first pass opened with who the user is and who pays. Theo said “These are actually good questions.” They often disagreed with the suggested answers, and found that useful.
- A developer on Hacker News got tired of plan mode. Grill-me asks until each design choice is set, then writes a PRD. It is now their go-to planning step.
- A developer on Hacker News sees value in it. But the plan sometimes “becomes overly prescriptive”, and agents chase goals that do not matter.
- GitHub issue 1145 shows a worse failure. A user said no to the suggested answer: “It documents what it recommended anyway.”
- A developer on Hacker News found that in Claude Code “Hi, /grill-me on X” does not work. The command must come first.
Skip it if your work is quick one-line fixes. An interview costs more than the task. Pick one install per agent, since both routes give each skill twice.
Superpowers
Price | Free, MIT license |
Install | One plugin command per agent |
Platforms | Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot CLI, OpenCode and more |
Best for | Multi-file features that need a spec first |
The problem. You ask for a feature that touches five files, and the agent starts editing before anyone agreed what it is. Three files in, you are reading its code to work out what it decided.
What it does. Superpowers makes the agent write a spec first. Ask it to build the feature, and it does not touch code. It asks what you are really trying to do and shows the spec in short chunks. Then it writes a plan and runs sub-agents once you say go. Small clear requests such as “make the icon cornflower blue” just get done. Bigger asks get a design doc and your approval first. Install it once per agent. Claude Code uses /plugin install superpowers@claude-plugins-official, and Cursor uses /add-plugin superpowers. Codex has it in its plugin marketplace.
What people found
- Evan Schwartz found plan mode made “a giant plan document”. Superpowers starts from a short sketch that is easier to judge. “I cannot recommend it enough.”
- The Gray Cat built a content pipeline they had put off for a year in about 3 hours. The planning phase was the best part. The downside: the check step wiped files it had just made.
- A developer on Hacker News settled on the plugin “across Claude and Codex, Cursor”. They pass the design doc between Claude and Codex, build with either, then review again. It is a lot of work, but the result beats doing it alone.
- A developer on Hacker News used it for a few weeks and wants an off switch. It checks each small edit, so a one-line change can take an hour.
Skip it if your work is small fixes in one file. The spec and plan cost more than the change.
Karpathy Guidelines
Price | Free, MIT license |
Install | Claude Code plugin, Cursor rule, or one CLAUDE.md file |
Platforms | Claude Code, Cursor, and any agent that reads an instructions file |
Best for | A light first step for any agent |
The problem. You ask for an export button and the agent ships a settings page, a service layer and a new config file. Nothing is broken. None of it is what you asked for.
What it does. Karpathy Guidelines is one file with four rules: think first, keep it simple, change only what you must, and work to a clear goal. The README quotes Andrej Karpathy: models “make wrong assumptions on your behalf and just run along with them”. One line sums it up: if 200 lines could be 50, rewrite it. Say “add a feature to export user data”. A bare agent guesses the users, the file and the fields. With the guidelines it asks first: all users or some, a download or a background job. Install it as a Claude Code plugin, drop CLAUDE.md into a project, or use its Cursor rule. One caveat from a developer on Hacker News: Karpathy did not write or endorse the file.
What people found
- A developer on Hacker News runs it everywhere: “I use it universally with all models and all harness.”
- A developer on Hacker News was asked how to improve agent code. One tip: “consider using something like Karpathy guidelines”.
- Dupple rates it for most people on Claude Code. But it is guidance, so the model can still ignore it.
- A developer on Hacker News calls the whole idea “quite ridiculous”. Vague rules will not make a solid process.
Skip it if you want rules enforced. Four rules in a file are a nudge, so keep your own review step.
Addy Osmani’s Agent Skills
Price | Free, MIT license |
Install | npx skills add addyosmani/agent-skills |
Platforms | Claude Code, Cursor, Codex, Copilot, Gemini CLI, OpenCode and more |
Best for | Shipping with spec, test and review steps |
The problem. The agent finishes the feature and marks it done. No tests, no second look at the code, and the gap only shows up after you ship.
What it does. Addy Osmani’s Agent Skills is 25 SKILL.md files for the whole build. Commands run from /spec and /plan to /test, /review and /ship. Install the lot with npx skills add addyosmani/agent-skills, or one file at a time with –skill. Type /test and the TDD skill makes the agent write the failing test first and watch it fail. Osmani’s own post explains the trick behind each file: it pairs the agent’s excuse with a rebuttal. For “I’ll write tests later”, the answer is “Later is the load-bearing word”. In Claude Code, call the review step /agent-skills:review, because /review is a built-in name. Gemini CLI takes gemini skills install with –path skills.
What people found
- A developer on Hacker News adopted a couple of them and says “the api design and ui testing ones have been particularly helpful.”
- A developer on Hacker News is “really impressed so far” on a side project. The pack holds their hand, so they can think about product design.
- A developer on Hacker News will take some skills but not the plugin. They worry about deleting it later, and think each skill should be customised to the developer.
Skip it if your agent tests and reviews its own work. If you only need the review or test step, take that one skill with –skill.
Impeccable
Price | Free, Apache-2.0 license |
Install | npx impeccable install |
Platforms | Claude Code, Cursor, Codex CLI, GitHub Copilot, Gemini CLI, OpenCode and more |
Best for | Reviewing and fixing pages you already have |
The problem. The page works. It also looks like other AI-built pages, with the same safe type.
What it does. Impeccable finds flaws in the pages you have. Run npx impeccable install, then /impeccable init once per project ($impeccable init on Codex). It reads your code, asks who the product is for, then writes PRODUCT.md. Then name a command and a target, such as /impeccable critique the pricing cards. It scores them against design rules and ranks what to fix first. The audit step ranks problems from P0 to P3. npx impeccable detect src/ runs rule checks with no LLM.
What people found
- A developer on Hacker News runs the scanner on their own sites: “It catches the actual visual tells.” They name dark glows on text and Inter as the default font.
- A developer on Hacker News found it “helpful for critiquing designs, developing a DESIGN.md guide” and iterating on UIs.
- A developer on Hacker News uses it inside Codex to make up for weak UI taste. Their verdict: it “still does not quite match the polish and instinct” Claude Code once had.
- A developer on Hacker News has used it and many others. They help with small fixes, but “the design itself is the same unless you are extremely explicit.”
- GitHub issue 744 logs an outside test. The skill fired 28 of 28 times when the model reached for it. Asked to write down the current visual system, it fired 0 of 10 times.
Skip it if the page does not exist yet, since the README says it is strongest on pages that exist. Also skip it if you are strict about what runs on your machine. It downloads an engine binary into ~/.impeccable/bin/. Telemetry stays on until you run npx impeccable telemetry off. Live mode can run the codex or claude CLI with approvals off unless you set IMPECCABLE_LIVE_COPY_AGENT.
Frontend Design
Price | Free |
Install | npx skills add, or the Claude Code plugin |
Platforms | Claude Code, Claude apps, other agents via npx skills |
Best for | Building a new page from a blank brief |
The problem. You ask for a settings page, then a pricing page, and both come back in the same skin. The brief barely changes what the agent reaches for.
What it does. Frontend Design is Anthropic’s own skill, rewritten on 3 September 2026. Ask it for a new page and it stops before code. You get a short plan with 4 to 6 named colours, typefaces and a layout. Then it checks that plan against your brief. It also names the looks AI pages fall into, such as warm cream with a terracotta accent. Any agent the skills CLI supports takes npx skills add https://github.com/anthropics/skills –skill frontend-design. Claude Code also has it as a plugin.
What people found
- A developer on Hacker News asks Claude for about 5 designs with it. They pick what they like, then ask for a new set.
- A developer on Hacker News says “even adding a simple frontend design skill improves the output substantially”.
- A developer on Hacker News gets better results by pointing Claude at a design system site.
- A developer on Hacker News calls it “good but only gets you so far”. It still needs a ton of steering.
- A developer on Hacker News keeps seeing a mix of sans serif and serif italics in its titles, as of 28 October 2026.
Skip it if the page exists and just needs a better look. Impeccable fits that job better. Skip it too if you have a design system you can point the agent at. More picks sit in the Claude design skills list.
Which Agent Skill To Install First
Start with Grill Me. It is the smallest change in this list, and the praise for it is the most consistent. It works in Claude Code, Codex, GitHub Copilot and Gemini CLI. Cursor and OpenCode take it through npx skills. Install the grilling skill with it, and in Claude Code keep /grill-me at the front of the prompt. The Matt Pocock skills roundup covers the rest of that pack, and adding skills to Cursor walks through that install route.
Reach for Superpowers only when features get bigger than one file. For more packs to browse, the Feedough Claude skills library collects each skill from these lists.
