innernet · coding
the best ai skills for coding
the skills that change how an agent writes code, not just what it knows. chosen for method, grouped by the job.
Most coding agents already know the syntax. What they lack is method: when to stop and plan, how to find a bug instead of guessing at it, what "done" means in your stack. That is what a good coding skill adds, and it is why the best ones are short documents about process rather than long ones about APIs.
The list below is grouped by the job each skill does. Every one is kept on innernet's community with its author and source, and every link is a page you can read first and then paste into your agent.
planning before building
brainstorming by obra. Makes the agent talk through intent and design before touching code, and wait for agreement. The single most effective fix for an agent that builds the wrong thing quickly.
writing plans and executing plans by obra. A pair: one turns an agreed design into small, checkable steps; the other works through them without drifting. Worth it for anything longer than an afternoon.
finding bugs
systematic debugging by obra. Root cause before fix: reproduce, trace, form a hypothesis, test it. It removes the habit of changing three things and seeing what sticks, which is how agents turn one bug into four.
tests and proof
test-driven development by obra. Red, green, refactor, enforced. Useful even if you don't practise TDD yourself, because it makes the agent prove behaviour instead of asserting it.
verification before completion by obra. The agent may not say "done" until it has run the thing and shown the output. Small, and it changes the whole texture of working with an agent.
webapp testing by Anthropic. Drives a real browser to check a web app actually works, rather than trusting that it compiles.
review and workflow
requesting code review by obra. How to hand work to a reviewer, human or agent, with the context they need.
using git worktrees by obra. Isolated working copies for parallel work, so an agent's experiment never lands on your branch.
frameworks, from the people who built them
react best practices and composition patterns by Vercel. Current React guidance from the team behind Next.js.
postgres best practices by Supabase. Schema, indexes, row-level security and the query mistakes that only show up under load.
cloudflare workers by Cloudflare. Writing Workers the way the runtime expects.
expo router by Expo, and react native by Vercel, for mobile.
claude api by Anthropic, for anything that calls a model.
interfaces
frontend design by Anthropic and web interface guidelines by Vercel. Two different takes on keeping generated UI from looking generated. Read both and keep the one that matches your taste.
building tools for agents
mcp builder and skill creator by Anthropic. For when you want to give an agent a new capability rather than use one.
how to choose
Install fewer than you think. Every skill's description sits in the agent's context, and five skills that match how you work beat forty that sort of do. A reasonable start for most developers is brainstorming, systematic debugging, verification before completion, and the one framework skill for your main stack.
Read each before you install it. They are short, and a skill is instructions your agent will follow with your repository and your terminal.
And keep them in one place. An agent in your terminal, another in your editor and a chat assistant on the side should all be following the same method. Using one skill in every ai tool covers how.
questions
what are the best claude code skills?
For method: brainstorming, writing and executing plans, systematic debugging, test-driven development and verification before completion, all from obra's superpowers collection. For frameworks: the skills published by the makers of your stack, such as Vercel for React, Supabase for Postgres and Cloudflare for Workers.
how many skills should a coding agent have?
Fewer than you think. Start with three or four that change how the agent works, plus one for your main framework, and add more only when you notice the agent repeatedly getting a job wrong.
do coding skills work in cursor or codex?
Skills are plain markdown, and several agents besides Claude Code read skill folders. Where a tool doesn't, the same skill can be handed over as a link or through a memory both tools are connected to.
where do i get these skills?
Each one links to its page on innernet's community, which names the author and links the original source. You can read it there, paste the link into your agent, or fetch the raw markdown by adding /raw to the link.
innernet is user-owned memory across AI tools, reachable over MCP. Your code decisions and the skills you work by, in one place, read by your terminal, your editor and your assistant alike. innernet.live
Checked 17 september 2026.
keep reading
one memory, every ai tool.
innernet keeps your projects, your decisions and your skills in one place, and every ai tool you connect reads the same thing. the skills people keep there are open to anyone.