🛠️ 66 Claude Code Skills Dropped This Week

66 skills. 817 security mappings. A Karpathy-inspired research loop that runs until it finds something worth keeping.

Today's Drops are heavy on Claude Code. That is not an accident. The community is converging on a pattern: stop treating Claude Code as a chat assistant and start treating it as a runtime. You give it a skill, a target, and a loop condition. It runs. You come back to results.

I have been watching this shift happen slowly for months. This week it broke into the open. Three separate Claude Code repos landed in the top 10 GitHub stars. One builds full-stack developer skills at scale. One automates your customer support flows. One turns autonomous research from a manual process into a self-correcting loop.

The interesting signal is not any single repo. It is that operators are now packaging their workflows as skills and sharing them. The skill layer is becoming the new plugin ecosystem. That matters for everyone building here.

Claude Tag also shipped this week, and it is worth reading carefully. An always-on AI teammate learning your Slack context is a productivity story on the surface. One layer down, it is a database of your company's institutional knowledge living on Anthropic's infrastructure. Your call on the tradeoff.

Today: 9 drops, 2 stack picks, 4 signals, one technique for running MCP servers without breaking your project config, and a builder move you can steal before lunch.

[Repo] ZhuLinsen/daily_stock_analysis, LLM-powered multi-market stock analysis system with multi-source market data, real-time news, a decision dashboard, and automated notifications. 46,919 stars on GitHub Trending. It runs on a zero-cost scheduled cadence, which means you can wire this to any brokerage data feed and have a morning briefing ready before you open your eyes.

[Skill] Jeffallan/claude-skills, 66 specialized skills for full-stack developers that turn Claude Code into an expert pair programmer. 10,151 stars. If you have been building skills one at a time, this is the shortcut: a vetted library you can pull from instead of starting from scratch.

[Skill] uditgoenka/autoresearch, A Claude Code autonomous research skill inspired by Karpathy's autoresearch pattern. 5,152 stars. The loop: Modify, Verify, Keep or Discard, Repeat. You set the goal, it runs until convergence. The gotcha is cost: uncapped loops will surprise you on a long session, so set a hard iteration ceiling before you run it.

[Repo] mukul975/Anthropic-Cybersecurity-Skills, 817 structured cybersecurity skills for AI agents, mapped across MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF, and MITRE F3. 19,527 stars. If you are building any agent that touches sensitive data or external APIs, this is the threat model you should be mapping against before you ship.

[Skill] nbashaw/claude-cs, A Claude Code skill for building custom customer support automation for your product. 90 stars but specific and deployable. The gap most operators hit is context: this skill is designed to ingest your product's specific flows, not generic helpdesk patterns.

[Repo] greggh/claude-code.nvim, Native Neovim integration for Claude Code. 2,077 stars. If your editor is Neovim and you have been context-switching out to the terminal to run Claude, this closes that gap: the assistant runs inside your buffer.

[Repo] ZhangHanDong/claude-code-api-rs, A high-performance Rust implementation of an OpenAI-compatible API gateway for Claude Code CLI. 170 stars. Niche but useful: if you are routing multiple tools through an OpenAI-compatible interface and want Claude Code in that pipeline, this is the bridge.

[Repo] pydantic/pydantic-ai, AI agent framework built the Pydantic way. 17,938 stars. Typed inputs and outputs for your agent calls, with the same validation model you already know if you have shipped anything with Pydantic. The practical gain: structured agent responses that don't surprise you downstream.

[Repo] CopilotKit/CopilotKit, Frontend stack for building agents and generative UI across React, Angular, mobile, and Slack. 35,429 stars. Makers of the AG-UI Protocol. If you need an agent-facing UI layer that is not a chat box bolted onto a dashboard, this is where to start.

[Tool] continuedev/continue, Open-source coding agent with 34,327 stars. Continue drops into VS Code and JetBrains as a sidebar assistant that knows your codebase, not just the file you have open. The non-obvious config move: point it at a local model via its config.json and you get full codebase awareness at zero API cost for the bulk of the work, then route only the hard reasoning calls to a frontier model.

[Tool] Jenqyang/Awesome-AI-Agents, A curated collection of autonomous LLM-powered agents. 1,167 stars. This is a reference layer, not a runtime: use it when you are scoping a new agent build and want to check whether the pattern already exists before writing from scratch. The discipline is searching before building.

Want to run a setup like this yourself? aigent-OS ($49) →

Claude Tag is in your Slack, learning your org. Anthropic's Claude Tag brings an always-on AI teammate directly into Slack. Beyond the productivity surface, it is a strategic capture of your company's institutional knowledge and context over time. Worth reading the data terms before you enable it org-wide. (TechCrunch)

$27 million spent on one local election to shape AI policy. Corporate AI super PACs dropped $27M on a New York 12th District race. This is the political-spending model for AI governance fights going forward. The labs are not waiting for Washington; they are buying the votes. (The Verge)

GPT-5 cracked a 3-year immunology mystery. A real-world case: GPT-5 Pro helped immunologist Derya Unutmaz find insights into T cell behavior that had stalled for three years. Useful data point if you are building any research-automation pipeline and need to explain the category to a skeptical client. (OpenAI Blog)

Simon Willison shipped an OPFS + Pyodide test harness. Willison is testing whether Datasette Lite, a Python app running entirely in-browser via Pyodide and WebAssembly, can edit persistent files using the Origin Private File System. If it ships cleanly, it closes a real gap for in-browser data tools that need write access without a server. (Simon Willison)

This week's technique: MCP servers. Wire a real tool into Claude Code so it acts, not just talks.

MCP (Model Context Protocol) servers give Claude Code direct access to external systems: GitHub, Supabase, Postgres, a browser, or anything else that has an MCP server written for it. The difference between Claude answering a question and Claude opening a PR is usually an MCP server.

1. Add the server to your project config. Create .mcp.json at the repo root (project scope, lives in version control) or add the mcpServers key to ~/.claude.json (user scope, available across all your projects). There is no .claude/mcp_settings.json, that path does not exist. 2. Define the server entry: { "mcpServers": { "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_TOKEN": "your_token" } } } }. Swap the package and env vars for whatever server you are adding. 3. Confirm the connection inside a session by running /mcp. Claude lists every connected server. If your server is missing, the JSON has a syntax error or the command path is wrong. Fix there first, not in the prompt.

You will know it worked when Claude stops saying "I can't access that" and starts doing the thing you pointed it at.

Claude Code docs: MCP

The move: batch structured extraction with a single Claude Code skill.

Most operators hit this wall: you have a spreadsheet, a CSV, or a directory of files and you need the same fields pulled from each one. The naive approach is to paste them one at a time. The operator approach is a skill that loops.

1. Create a skill at .claude/skills/extract.md. In the skill, define the target schema: the exact fields you want, the output format (JSON), and any edge-case rules (what to do when a field is missing). 2. Invoke it with the file list: /extract followed by a glob or a path. The skill reads each file, extracts to schema, and writes structured output. No manual iteration. 3. Add a validation step in the skill: after extraction, compare required fields to the schema and flag any row that does not conform. Claude surfaces gaps instead of silently dropping them.

You will know it worked when your output file has zero blank required fields and a logged list of rows that need human review. That list is the product, not the extracted data.

How does Claude Code get better the more you use it? aigent-OS is a self-improving Operator OS built specifically for Claude Code: it learns your patterns, compounds your best moves, and makes the whole system sharper over time. At $49 it is the kind of infrastructure decision that pays for itself the first week. Check it out in the store via the button below.

Every issue this week has pointed at the same thing: the operators pulling ahead are the ones treating Claude Code as a runtime with a skill layer, not a smarter search box. The repos are there. The skills are shipping. The distance between reading this and running it is about thirty minutes.

See you Thursday.