Best AI Coding Tools on GitHub for .NET Devs

I spent the last few weeks researching agent skills that are relevant and useful for professional .NET developers working with AI-assisted tools, specifically Claude Code. I’ve listed them here but in case you’re new to skills I’ll start with some key things to know first.

What an “agent skill” actually is

A skill is a folder with a SKILL.md file at the top, plus optional scripts and reference docs underneath. The SKILL.md has YAML frontmatter (just name and description are required) and a markdown body the agent reads when it decides the skill is relevant.

The standardisation of SKILL.md is one of those rare cases where a single company’s feature became an industry-wide standard in a matter of weeks rather than years. Anthropic published the Agent Skills specification at agentskills.io, releasing it as an open standard under Apache 2.0 (code) and CC-BY-4.0 (docs), with a reference SDK and validation tooling. Within days of publication, Microsoft had integrated SKILL.md support into VS Code and GitHub Copilot. OpenAI adopted a structurally identical architecture in Codex CLI and ChatGPT. Cursor and others followed.

Skills can be installed through a few different paths: via marketplace plugins (like Anthropic’s official marketplace or community ones such as VoltAgent’s awesome-agent-skills), via CLI installers (e.g., npx skills add <org>/<repo> or openskills, the universal installer for SKILL.md - npx openskills install anthropics/skills), or manually by cloning a repo and dropping the skill folder into the right location.

So when I install one of these repos, the same SKILL.md is portable across tools that support the Agent Skills open standard, including Codex and Cursor - and that’s usually handled by the installation process depending on which agent tool you’re using. If you manually install a skill for Claude Code, it needs to live in .claude/skills/ (project-level) or ~/.claude/skills/ (personal), or be symlinked from a shared location into both .claude/skills/ and Codex’s .agents/skills/ (and .cursor/skills/ for Cursor) to avoid duplicating the files on disk.

A few things to keep in mind before you go install any of these:

• Don’t confuse a skill with a prompt. A prompt is one-off context. A skill loads on demand and persists across sessions.
• The description field is the most important thing in the file. If your skill never triggers, it’s almost always because the description doesn’t tell the model when to use it.

OK. The repos.

1. forrestchang/andrej-karpathy-skills: LLM coding discipline in one file

A single CLAUDE.md or Cursor rule built from Andrej Karpathy’s observations on why LLMs fail at coding tasks. Not a prompt. Not a system message. A behavioral contract you drop into any repo.

Four principles, each targeting a specific failure mode:

Goal-Driven Execution is the one that actually changes agentic behavior. Instead of “fix the bug,” it effectively says “write a test that reproduces the bug, then make it pass.” The model loops on a verifiable success condition instead of guessing what done looks like.

The repo also ships a committed Cursor rule (.cursor/rules/karpathy-guidelines.mdc) so the same guidelines apply across both tools without duplicating anything.

2. microsoft/azure-skills: Real Azure work, not generic cloud advice

Claude Code · Cursor · Codex

Microsoft’s official agent plugin for Azure. Three layers in one install:

Without the MCP layer, the agent talks confidently about Azure and produces plausible-sounding resource names. With it, it calls az directly and returns real tool-backed responses from your actual subscription.

Skills cover the full Azure workflow surface: azure-prepare, azure-validate, azure-deploy, azure-upgrade, azure-diagnostics, appinsights-instrumentation, azure-cost, azure-compute, azure-rbac, entra-app-registration, azure-ai, azure-messaging, azure-storage, azure-kubernetes, and more.

Then browse plugins using /plugins and install azure.

3. dotnet/skills: The .NET team’s official agent skills

Claude Code · Cursor · Codex

The .NET team’s curated marketplace of agent skills, organized into focused plugins that map to how .NET work actually gets done. Install only what you need.

What separates this from community .NET prompt packs is that the team actually measures whether the skills work. Each skill ships with its own eval in the repository that you can inspect and run. There is a public accuracy dashboard showing scoring trends over time.

These skills come from the team that ships the platform. They are the same workflows used internally and tested against real engineering scenarios, not generic prompts written against the docs.

Who skips it: anyone not writing C# or F#.

4. github/awesome-copilot — The community marketplace for Copilot

GitHub’s official community-curated collection of agents, instructions, skills, plugins, hooks, agentic workflows, and API recipes for Copilot. The scope is wide: everything from a New Relic incident-response agent to a tool-guardian hook that blocks dangerous commands before they execute.

The awesome-copilot marketplace is registered in Copilot CLI by default, so no setup is needed. Browse the full collection with search and filtering at awesome-copilot.github.com, which also includes a Tools section for MCP servers and a Learning Hub for tutorials.

Quality varies because it is community-contributed. Treat it like npm: useful, but read the source before installing anything in a production workflow.

I’m phasing out use of Github Copilot in favour of Claude Code, Codex and Cursor but these skills are reusable and can just be copied cover into agent skills for Claude Code or Cursor. The marketplace itself is Copilot-specific, but the underlying SKILL.md files are portable across tools.

5. obra/superpowers : A complete agentic development workflow

Where most skills fix one specific failure mode, Superpowers installs an entire software development methodology. The difference is scope: this is not a behavioral nudge, it is a structured workflow the agent follows from first conversation to merged branch.

The core loop:

The skills trigger automatically. The agent checks for relevant skills before any task. These are mandatory workflows, not suggestions.

Beyond the core loop, the library also covers systematic-debugging (4-phase root cause process), dispatching-parallel-agents (concurrent subagent workflows), receiving-code-review, and writing-skills for extending the system itself.

The philosophy maps directly to what most .NET codebases need more of: TDD enforced at the agent level rather than aspired to in a wiki, YAGNI as a hard constraint rather than a code review comment, and verification before the agent declares anything done.

Built by Jesse Vincent and the team at Prime Radiant. Read the origin post: Superpowers for Claude Code.

6. JuliusBrussee/caveman: Cut ~75% of your output tokens

A skill that makes the agent drop articles, filler, hedging, and pleasantries while keeping all technical substance. Based on a viral observation that caveman-speak dramatically reduces LLM token usage without losing accuracy. Someone turned it into a one-line install.

Real token counts measured against the Claude API:

A March 2026 paper, Brevity Constraints Reverse Performance Hierarchies in Language Models, found that constraining models to brief responses improved accuracy by 26 percentage points on certain benchmarks. Shorter output is not just cheaper, it is often more correct.

Four intensity levels: lite (drop filler, keep grammar), full (default caveman), ultra (telegraphic), and wenyan (classical Chinese, genuinely the most token-efficient written language ever invented).

The bonus tools are where the real leverage is for .NET devs:

• caveman-commit: terse Conventional Commits, max 50 char subject, why over what
• caveman-review: one-line PR comments: L42: bug: user null. Add guard.
• caveman-compress: rewrites your CLAUDE.md into compressed form, cutting roughly 46% of input tokens on every session start. The original is preserved as CLAUDE.md.original for humans.

That last one compounds. Your CLAUDE.md loads on every single Claude Code session. Cutting it by half cuts every session startup cost forever.

One caveat: caveman only affects output tokens. Thinking and reasoning tokens are untouched. The biggest practical win is readability and response speed, not just cost.

7. nextlevelbuilder/ui-ux-pro-max-skill: Design intelligence for your agent

Tackles the problem agents have with UI work: technically correct output that looks average compared to human designers. When you ask for a landing page, the agent needs to know which color palette suits a fintech app versus a wellness brand, which typography pairing fits the tone, and which patterns to avoid entirely.

This skill bundles a design-system reasoning engine. When you describe a UI task, it runs five parallel searches across:

• 161 industry-specific reasoning rules (SaaS, fintech, healthcare, beauty, gaming, Web3, and more)
• 67 UI styles (Glassmorphism, Brutalism, Bento Grid, AI-Native UI, Soft UI, Cyberpunk, and more)
• 161 color palettes matched 1:1 with product types
• 57 font pairings with Google Fonts imports already wired
• 24 landing-page patterns and 25 chart type recommendations

The output is a complete design system: pattern, style, colors, typography, effects, anti-patterns to avoid, and a pre-delivery checklist covering contrast ratios, focus states, cursor behavior on interactive elements, and responsive breakpoints at 375, 768, 1024, and 1440px.

The anti-pattern catalog is particularly useful. It explicitly tells the agent things like “no AI purple/pink gradients on a banking app” - the kind of taste rule LLMs violate by default without supervision.

Supports 15 stacks: React, Next.js, Vue, Nuxt, Astro, Svelte, SwiftUI, React Native, Flutter, Angular, Laravel, Jetpack Compose, HTML + Tailwind, shadcn/ui, and Nuxt UI. Mention your stack in the prompt or it defaults to HTML + Tailwind.

How to stack these

Not all seven on every project. The order that makes sense:

1. andrej-karpathy-skills on every project, before anything else. It is the discipline layer that stops the agent overengineering before the first task runs.
2. microsoft/azure-skills the moment the project has an azure.yaml or a Bicep file.
3. dotnet/skills on every .NET project. Install only the plugins matching the work: dotnet, dotnet-msbuild, dotnet-aspnet, dotnet-data as needed.
4. awesome-copilot any .NET project. Take .NET ad hoc related skills when a specific need comes up.
5. superpowers on any project where you are running multi-step autonomous tasks. It is the workflow layer: spec, plan, TDD, subagent execution, review, and merge as a single enforced loop. Not intended to use for POC projects.
6. caveman on any project. The input token savings compound across every session.
7. ui-ux-pro-max-skill only on projects with a UI and no existing design system.

Happy building.