GitHub Weekly Wins: 13 Repos Reshaping AI Agents

> **TL;DR:** This week's GitHub roundup is dominated by 'agent skills' repositories — reusable, composable instructions that make coding agents like Claude Code behave more like disciplined engineers. Alongside those, a token-compression library, an open-source CapCut alternative, and a security scanner for AI skills round out a week focused squarely on making agents cheaper, safer, and more reliable to run.

Key Takeaways

- mattpocock/skills leads the week with over 159,000 stars, packaging one engineer's personal Claude Code workflow into reusable skills - Agent 'skills' repositories dominate the list — engineering, marketing, product management, and research skills are all being open-sourced and shared - headroomlabs-ai/headroom tackles a real cost problem: cutting 60-95% of the tokens agents waste on raw tool output and logs - NVIDIA's SkillSpector addresses a growing risk — the project cites research finding 26.1% of installed agent skills contain vulnerabilities - OpenCut and apple/container show the open-source alternative trend extending beyond AI into video editing and containerization

Why This Week's List Looks Different

Most weeks on [GitHub Weekly Wins](https://speka.info/github-weekly-wins/), the trending list is a grab-bag of frameworks, dev tools, and the occasional meme repo. This week it isn't. Of the thirteen repositories that broke out, more than half are "agent skills" — structured instruction sets that teach coding agents like Claude Code how to work the way a senior engineer, marketer, or product manager would. The rest cluster around a second theme: making those same agents cheaper to run and safer to trust.

Here's the full list, ranked by GitHub stars.


mattpocock/skills

**What it is:** A public dump of the personal Claude Code skills used daily by developer-educator Matt Pocock, covering real engineering workflows rather than "vibe coding" shortcuts. The repo explicitly positions itself against heavier process frameworks like GSD, BMAD, and Spec Kit, arguing that owning your whole workflow makes bugs harder to isolate.

**Why it matters:** With close to 160,000 stars, it's the most-starred repo in this week's list by a wide margin, and it doubles as a live case study in how individual developers are now open-sourcing their agent configurations the way they once open-sourced dotfiles.

**Who it's for:** Developers already using Claude Code who want small, composable, model-agnostic skills instead of an all-or-nothing agentic framework.

**GitHub:** [mattpocock/skills](https://github.com/mattpocock/skills)

iptv-org/iptv

**What it is:** A community-maintained collection of publicly available IPTV channel playlists from around the world, distributed as M3U files alongside an EPG (electronic program guide) and API.

**Why it matters:** At over 130,000 stars, it's a reminder that not every trending repo is AI-related — this one's utility is simply pasting a link into any streaming-capable video player and getting live TV.

**Who it's for:** Anyone building or using a media player that supports live streaming, and hobbyists who want a single, continuously updated source of public broadcast links.

**GitHub:** [iptv-org/iptv](https://github.com/iptv-org/iptv)

addyosmani/agent-skills

**What it is:** A set of production-grade engineering skills for AI coding agents, built around eight slash commands — `/spec`, `/plan`, `/build`, `/test`, `/review`, `/webperf`, `/code-simplify`, and `/ship` — that map directly onto the software development lifecycle.

**Why it matters:** Rather than a single mega-prompt, each command activates the right combination of skills automatically, encoding quality gates senior engineers already use so agents apply them consistently across Claude Code, Cursor, Codex, and Antigravity.

**Who it's for:** Teams standardizing how multiple coding agents move from spec to shipped code, especially those juggling more than one agent tool.

**GitHub:** [addyosmani/agent-skills](https://github.com/addyosmani/agent-skills)

OpenCut-app/OpenCut

**What it is:** A free, open-source video editor built as an alternative to CapCut, currently being rewritten from the ground up with a Rust core, a plugin-first architecture, and a planned MCP server for AI agent control.

**Why it matters:** Nearly 62,000 stars signal real demand for a self-hostable editor that isn't locked to one vendor's cloud — and the roadmap's headless/automation mode and scripting tab point toward editors agents can drive directly, not just humans.

**Who it's for:** Creators and developers who want a CapCut-style editor they can run themselves, plus anyone watching for the first AI-agent-native video editor to reach maturity.

**GitHub:** [OpenCut-app/OpenCut](https://github.com/OpenCut-app/OpenCut)

headroomlabs-ai/headroom

**What it is:** A context-compression layer that sits between AI agents and everything they read — tool outputs, logs, files, and RAG chunks — claiming 60-95% token reductions on structured data like JSON and 15-20% reductions for general coding-agent traffic, available as a library, a proxy, or an MCP server.

**Why it matters:** Token costs are quickly becoming the real constraint on agent usefulness, a tension we've already covered in [the debate GLM 5.2 sparked over AI margin collapse](https://speka.info/blog/glm-5-2-sparks-ai-margin-collapse-debate-on-hn). Headroom attacks that cost problem directly, and reversibly, without changing the underlying model. It's a similar efficiency instinct to what's driving lightweight, run-anywhere models like [Ternlight's 7MB browser embedding model](https://speka.info/blog/ternlight-7mb-embedding-model-runs-in-your-browser).

**Who it's for:** Anyone running agents against verbose tool output or large RAG pipelines who wants smaller context windows without rewriting prompts.

**GitHub:** [headroomlabs-ai/headroom](https://github.com/headroomlabs-ai/headroom)


Panniantong/Agent-Reach

**What it is:** A single CLI that gives AI agents read and search access to platforms that normally block scraping or require paid APIs — Twitter/X, Reddit, YouTube, GitHub, Bilibili, and XiaoHongShu — without per-platform API fees.

**Why it matters:** Agents are increasingly expected to research the live web, not just static docs, and this repo's pitch is that it absorbs the maintenance burden of each platform's anti-scraping changes so the agent's setup doesn't break every time a site updates its defenses.

**Who it's for:** Developers building research or monitoring agents that need to pull real-time signal from social and video platforms in one integration instead of many.

**GitHub:** [Panniantong/Agent-Reach](https://github.com/Panniantong/Agent-Reach)

asgeirtj/system_prompts_leaks

**What it is:** A regularly updated archive of extracted system prompts from major AI products, including Claude, ChatGPT, Gemini, Grok, Copilot, and coding tools like Cursor and Codex — documented for anyone studying how these assistants are actually instructed.

**Why it matters:** The project was featured by The Washington Post, which used it to illustrate the hidden rules shaping chatbot behavior, and it now tracks prompt diffs between model versions so builders can see exactly what changed release to release.

**Who it's for:** Prompt engineers, researchers, and builders who want a reference for how leading AI products structure their own instructions.

**GitHub:** [asgeirtj/system_prompts_leaks](https://github.com/asgeirtj/system_prompts_leaks)

mvanhorn/last30days-skill

**What it is:** An agent skill that researches any topic across Reddit, X, YouTube, Hacker News, Polymarket, and the wider web, then synthesizes a single grounded summary — describing itself as a search engine scored by upvotes, likes, and real money rather than editors.

**Why it matters:** It ships with Reddit, Hacker News, Polymarket, and GitHub working out of the box, with a one-time setup wizard unlocking X, YouTube, and TikTok — lowering the bar for giving an agent genuinely current, cross-platform research ability.

**Who it's for:** Anyone who wants an agent to answer "what's actually happening with X right now" using recency-weighted, multi-source evidence instead of a single search query.

**GitHub:** [mvanhorn/last30days-skill](https://github.com/mvanhorn/last30days-skill)

apple/container

**What it is:** Apple's own tool for creating and running Linux containers as lightweight virtual machines on a Mac, written in Swift and optimized specifically for Apple silicon, built on top of Apple's Containerization package.

**Why it matters:** It consumes and produces standard OCI-compatible images, so it can pull from and push to any normal container registry — giving Apple silicon users a native alternative to Docker Desktop without leaving Apple's own tooling.

**Who it's for:** Developers on M-series Macs who want container workflows without a third-party virtualization layer.

**GitHub:** [apple/container](https://github.com/apple/container)

coreyhaines31/marketingskills

**What it is:** A collection of AI agent skills focused specifically on marketing work — conversion rate optimization, copywriting, SEO, analytics, and growth engineering — built to work with Claude Code, Codex, Cursor, and Windsurf.

**Why it matters:** It extends the same "skills" pattern seen in engineering-focused repos this week into a completely different discipline, aimed at technical marketers and founders rather than developers.

**Who it's for:** Founders and marketers who already use a coding agent and want it to handle CRO audits, landing-page copy, and SEO work with the same rigor as code.

**GitHub:** [coreyhaines31/marketingskills](https://github.com/coreyhaines31/marketingskills)


phuryn/pm-skills

**What it is:** A marketplace of more than 100 agentic skills, commands, and plugins for product management, spanning discovery, strategy, execution, launch, and growth — including 68 individual PM skills and 42 chained workflows across nine plugins.

**Why it matters:** It packages product-management judgment calls — writing a PRD, planning a launch, defining metrics — into commands like `/discover`, `/strategy`, and `/plan-launch`, aiming to bring the same consistency to product decisions that engineering skills bring to code.

**Who it's for:** Product managers using Claude Code or similar tools who want structured, repeatable workflows instead of ad-hoc prompting.

**GitHub:** [phuryn/pm-skills](https://github.com/phuryn/pm-skills)

NVIDIA/SkillSpector

**What it is:** A security scanner built specifically for AI agent skills, designed to detect vulnerabilities, malicious patterns, and other security risks before a skill gets installed into an agent like Claude Code, Codex CLI, or Gemini CLI.

**Why it matters:** The project's own research is the sobering part: it reports that 26.1% of scanned skills contain vulnerabilities and 5.2% show signs of likely malicious intent — a real concern given how many of the skills in this very roundup ask users to drop shell scripts straight into their agent's directory. SkillSpector answers the question "is this skill safe to install?" before you find out the hard way.

**Who it's for:** Anyone installing third-party agent skills or plugins who wants a vetting step, and teams standardizing what's allowed to run inside their agents.

**GitHub:** [NVIDIA/SkillSpector](https://github.com/NVIDIA/SkillSpector)

music-assistant/server

**What it is:** A free, open-source media library manager that connects to streaming services and a wide range of connected speakers, designed to run continuously on an always-on device like a Raspberry Pi, NAS, or Intel NUC.

**Why it matters:** It's the smallest repo on this week's list by star count, but it fills a specific gap — a self-hosted hub that unifies music sources and smart speakers without depending on any single streaming vendor's app.

**Who it's for:** Home-lab and smart-home enthusiasts who want one server managing music across multiple streaming accounts and speaker brands.

**GitHub:** [music-assistant/server](https://github.com/music-assistant/server)

The Bigger Pattern

Line these thirteen repos up and a shape emerges: the ecosystem is professionalizing around AI coding agents from three directions at once. Skills repos (from mattpocock, Addy Osmani, Corey Haines, and phuryn) are turning tribal engineering and business knowledge into portable, shareable instructions. Infrastructure repos like headroom and Agent-Reach are solving the unglamorous plumbing problems — cost and access — that determine whether an agent can actually finish a task. And SkillSpector's existence is itself a signal: once skills spread this fast, someone has to start checking what's inside them, much like scrutiny is already building around frontier coding models such as the [rumored GPT-5.6 Sol Ultra tier for OpenAI Codex](https://speka.info/blog/gpt-5-6-sol-ultra-rumored-for-openai-codex). Expect next week's list to keep tilting the same direction — more skills, more compression, and more tooling built to keep the whole stack honest.

Frequently Asked Questions

What is an 'agent skill' on GitHub?

An agent skill is a packaged set of instructions, workflows, and quality checks that a coding agent like Claude Code, Codex, or Cursor can load to follow a consistent process — for example, writing a spec before code, or reviewing output before shipping.

Why did agent skills dominate this week's GitHub trending list?

Skills are cheap to write, easy to share as a repo, and work across multiple agent tools, so individual developers and companies are open-sourcing their internal workflows the way they once shared dotfiles or boilerplate templates.

What does headroomlabs-ai/headroom actually compress?

It compresses tool outputs, logs, files, and RAG chunks before they reach the LLM, claiming 60-95% token reductions on structured data like JSON and 15-20% reductions in general coding-agent workflows.

Is it safe to install AI agent skills from GitHub?

Not automatically — NVIDIA's SkillSpector project cites research finding that 26.1% of scanned agent skills contained vulnerabilities and 5.2% showed signs of likely malicious intent, which is why scanning tools like it are emerging alongside the skills themselves.

Is OpenCut a full replacement for CapCut today?

The actively maintained OpenCut Classic version is usable now, but the flagship OpenCut repo is mid-rewrite toward a Rust-core, plugin-first editor with planned MCP and headless support, so some announced features aren't shipped yet.

Sources

- https://github.com/mattpocock/skills - https://github.com/iptv-org/iptv - https://github.com/addyosmani/agent-skills - https://github.com/OpenCut-app/OpenCut - https://github.com/headroomlabs-ai/headroom - https://github.com/Panniantong/Agent-Reach - https://github.com/asgeirtj/system_prompts_leaks - https://github.com/mvanhorn/last30days-skill - https://github.com/apple/container - https://github.com/coreyhaines31/marketingskills - https://github.com/phuryn/pm-skills - https://github.com/NVIDIA/SkillSpector - https://github.com/music-assistant/server