🛠️ 345 Claude Code Skills in One Repo

Kimi K3 dropped last week and the conversation hasn't stopped. China's Moonshot AI released an open-weight model that Axios reports is benchmarking against the top closed labs, and it is not a rounding error. Frontier capability plus open weights is a real combination now, and it changes what you can self-host.

The CNBC piece on White House access controls is worth reading alongside it. CNBC reports the administration is moving to shape which frontier models get released and to whom, though the White House disputes that it approves releases. Closed labs under political access constraints versus open-weight models anyone can pull: that is the actual race shaping your stack decisions right now.

Meanwhile the Claude Code skill shelf kept filling up. This week's Drops run nine deep: four skill packs plus five repos, including the code-review-graph project that goes straight at the context-bloat problem every operator hits eventually. The full list is below.

One note before you get to it: if you are still feeding your entire codebase into every AI coding session, the code-review-graph repo in Drops will change your workflow today.

New here? This part's for you. (Operators, skip to The Drops.)

What is open source?

When a project is open source, the people who built it have shared all the code publicly, for free. Anyone can read it, copy it, modify it, and build on top of it. Think of it like a recipe being published in a cookbook versus kept secret in a restaurant kitchen. The cookbook version spreads; people improve it; better versions come back. Most of the repos in today's Drops are open source, which is why you can clone them and run them yourself instead of paying for a subscription.

Do this today

Open GitHub (github.com) and search for one thing you wished you had an AI tool to do. Browse the top results. You are not installing anything yet. You are just learning what the open-source community has already built. Five minutes of looking beats six months of "I should really build that."

How to use today's picks

Today's Drops include nine picks: five repos and four Claude Code skill packs. For each one, read the one-line pitch here, click through, and skim the README (the README is the project's front-page notes: what it does, how to start, what you need installed). If something solves a real problem you have right now, bookmark it. Installing comes later, only when you need it.

Plain English

Repo: Short for "repository." A folder on GitHub that holds a project's code, notes, and history. When someone says "clone the repo," they mean download that folder to your computer.

README: The front page of any repo. Usually a plain text file that explains what the project does, how to install it, and how to get started. Always read it before you try to run anything.

We read your replies

You told us you're here to learn, and that's exactly why this section exists. If there's a word from today's issue you want decoded in plain English, hit reply and send it over. We read every one.

[Skill] alirezarezvani/claude-skills, 345 Claude Code skills, 30+ agents, 70+ custom commands, and 330+ skill files across Claude Code, Codex, Gemini CLI, Cursor, and eight other tools. This is the skill pack I wish existed a year ago. Drop the ones you need into your .claude/skills/ folder and stop rewriting the same instructions from scratch every project.

[Repo] tirth8205/code-review-graph, 21,018 stars. Builds a persistent intelligence graph of your codebase so AI coding tools stop reading files they do not need. If your context window fills up on large repos before Claude finishes a review pass, this is the fix. MCP and CLI interfaces both ship with it.

[Repo] kvcache-ai/ktransformers, 18,298 stars. A flexible framework for heterogeneous LLM inference and fine-tune optimizations. If you are running models across mixed hardware (CPU offloading, different GPU tiers), this handles the routing. Less hand-tuning, more throughput.

[Repo] mozilla-ai/llamafile, 25,391 stars. Run an LLM as a single executable file. No Python environment, no Docker, no dependency hell. One file, double-click, model runs. The Social Buzz signal on this one has been building for weeks, and for good reason, it is the fastest path from zero to local inference.

[Skill] aiwithremy/claude-skills-llm-council, 1,167 stars. A Claude Code skill that routes a decision through five AI advisors with peer review before returning a verdict. Useful when you are evaluating an architectural choice and do not want to anchor on the first answer Claude gives.

[Skill] SnailSploit/Claude-Red, 2,716 stars. A curated library of offensive security skills for the Claude skills system. Each skill primes Claude with expert-level security context for a specific attack pattern. If you are running red-team reviews on your own infrastructure, this saves the prompt-engineering work.

[Skill] Owl-Listener/designer-skills, 1,863 stars. Agentic design skills covering research, design systems, UI, interaction, and delivery. If Claude Code is doing front-end work on your project, these skills tighten the output considerably. Less back-and-forth, more shipped components.

[Repo] alvinreal/awesome-opensource-ai, 4,170 stars. A daily-updated curated list of genuinely open-source AI projects, models, tools, and infrastructure. The word "truly" in the description matters: this one filters out the open-washing. Good bookmark when you are evaluating the landscape before committing to a stack.

[Repo] unslothai/unsloth, 2x faster fine-tuning on 60% less VRAM, for Llama, Mistral, and Gemma. If you have been putting off fine-tuning a model because your hardware felt too small, this is your practical entry point. The numbers are verified and the community around it is active.

[Affiliate] Runpod, GPU cloud for running open-weight models. Every model in today's issue has to run on a GPU your laptop probably lacks. Spin up an H100 or A100 pod billed by the second to try Kimi K3 or fine-tune with unsloth, then scale to zero when the job's done. We may earn a commission.

[MCP] herry2059/project-os-for-codex, An open-source control plane for Codex projects with Git-backed context, visible agent progress, scoped MCP access, and resumable work. The part that matters most is safe handoffs: when an agent run fails mid-task, you pick up where it stopped instead of starting over. If you run multi-step Codex sessions that sometimes brick out, this addresses the exact failure mode.

[MCP] JesusRS1/stock-trade-finance-api, Connects your MCP environment to Tiingo for real-time stock prices, news, forex, and company fundamentals. The non-obvious config move: use the fundamentals endpoint to pipe balance-sheet data into an analysis agent rather than just price feeds. Turns a price-ticker integration into an actual research layer.

China's open-weight insurgency is not slowing down. Kimi K3 from Moonshot AI benchmarks against frontier closed models and ships with open weights. Axios frames it as a full split in the AI race: closed-weight dominance versus open-weight disruption. For operators, the practical read is this: the cost floor for frontier-grade capability just dropped again.

White House is now a gatekeeper for frontier model releases. The Trump administration is actively dictating which models get released and who gets access, according to CNBC. If your stack depends on frontier API access from US labs, you now have a regulatory dependency you did not have six months ago. Open weights look more resilient against that risk.

Grok's CLI was uploading your local files. Security researchers caught xAI's Grok Build CLI uploading entire git repositories to xAI's cloud storage, not just the files it was asked to read. Not a Claude story, not a generic AI story: a specific tool, a specific behavior, a specific exposure. Audit what your AI coding tools are sending before you run them on client repos.

Vercel Sandbox egress is now free. Data downloaded by the Vercel Sandbox during builds (package installs, repo clones, external datasets) no longer counts toward your Sandbox Data Transfer bill. Small change on paper; meaningful if you are running CI-heavy agent workflows on Vercel's platform.

Open-weight inference is becoming its own control point. Forbes covers the shift: managed inference for open-weight models is consolidating, and the operators who control that layer are building durable positions. This is the infrastructure layer under your model choices, and it is moving fast.

Move: Wire a skill pack into a new project in under five minutes

The claude-skills repo (linked in The Drops) ships 345 pre-built skills. Most operators drop them into a project and stop there. The sharper move is to treat them as a starting library, not a final answer.

1. Clone the repo. Browse the /agents and /skills folders and pull only the files relevant to your project into your .claude/ directory. Do not install all 345. Context budget matters. 2. Open one skill file and read the structure. Note how the SKILL.md format primes Claude with domain context before any task. That pattern is the thing to steal. 3. For any skill that is close but not right, copy it, rename it, and edit the domain instructions. You now have a custom skill that took ten minutes instead of two hours. 4. Run a task using the skill and check whether Claude's first response is noticeably better than without it. That delta is your signal for whether the skill is earning its context cost.

You will know it worked when Claude stops asking clarifying questions that a competent domain expert would already know, and starts asking the questions that actually move the work forward.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. This week: a single period that quietly broke every link in a batch of outreach.

The smallest character in a sentence can be the one that costs you the click.

We staged a batch of outreach emails. Each one closed with a link to our media kit, and each link was followed, naturally, by a period. It read perfectly in the composer. It arrived broken. In plain text mail, that trailing period fuses to the end of the URL, and the mail client folds the whole string, dot included, into a dead link. Every email in the batch, the same silent break, in the exact spot we most wanted a click.

Nothing errored. The draft looked finished. The bug lived entirely in the gap between how the message is written and how it renders where it lands.

The fix is trivial once you see it: keep punctuation off the end of a URL, and put the link on its own line so nothing can touch it. The habit behind the fix is the real lesson. Preview your work in the place it will actually be read, not the place you wrote it. The composer is a flattering mirror. The inbox is the truth.

A rule that has saved us more than once: if it renders somewhere, test it there, not upstream of there.

What is the tiniest bug that has cost you the most? Hit reply and tell us. We read every one.

The open-weight insurgency and the skill-pack ecosystem are moving in the same direction: the operator who moves fast on both ends up with capabilities the tool-dependent competitor is still waiting on a waitlist for. That is the position worth building toward.

See you Tuesday.