GitHub Trending Top 10: Skills become the workbench

GitHub's weekly chart is unusually concentrated this time: four of the ten entries are installable skill packs, while the rest turn agents into video editors, trading systems, tutors, Office operators, UI critics, or terminal workers. The top two also show the split in maturity. Hallmark is a compact design skill with no published release; OpenCut has 1,584 commits but is rebuilding its editor from scratch. 1

Coverage: Jul 12, 2026 at 20:00 through Jul 19, 2026 at 20:00 Eastern Time. The ranking below is the weekly GitHub Trending page captured for this run. Star totals and weekly gains are snapshots from that page, not a claim that GitHub exposes a permanently auditable seven-day series.

Hallmark attacks a familiar failure mode in AI-built interfaces: the page works, but it looks like the statistical average of every landing page in the model's training data. Its design skill for Claude Code, Cursor, and Codex chooses a macrostructure for a brief, applies one of twenty themes, runs fifty-seven anti-slop gates, and performs a self-critique before emitting a page. The output is self-contained HTML and CSS, which keeps the result inspectable rather than hiding the design logic in a hosted generator. 2

The interesting design decision is to make variation structural. hallmark audit scores existing code without editing it, redesign keeps the copy and information architecture while replacing the visual fingerprint, and study extracts design DNA without producing a pixel clone. That is a sharper distinction from a component library or a prompt pack, both of which usually improve consistency at the cost of sameness. The trade-off is maturity: GitHub lists 133 commits, 20 open issues, MIT licensing, and no published releases. 3 Star it if you want a small, inspectable example of quality control inside an agent skill; skip it as a production design system until the release story catches up with the star surge.

OpenCut is the week's biggest growth event by stars, but its README contains an important qualifier: the project is being rewritten from the ground up. It is positioned as a free, open-source CapCut alternative for web, desktop, and mobile. The rewrite is planned around an Editor API, first-class third-party plugins, one codebase with a Rust core, an MCP server for agents, headless batch rendering, and an in-editor scripting tab. Those pieces point to an editor that can become programmable infrastructure instead of only a manual timeline UI. 4

That ambition is also the main risk. The repository is currently 97.2% TypeScript with only 0.7% Rust, while the proposed cross-platform core is still part of the future architecture. The classic version remains at opencut-app/opencut-classic, and the public site still runs that version. The GitHub API shows 1,584 commits, 355 open issues, three releases, and MIT licensing; the latest listed release is v0.3.0 from Apr 15. 5 Star it as a bet on an open, agent-ready media stack; skip using the rewrite as your editor until the new architecture becomes the shipped product.

This is less a library than a runnable catalog. Awesome LLM Apps groups more than 100 agents, skills, RAG applications, voice systems, MCP integrations, always-on workers, generative interfaces, and framework crash courses. The repository's useful unit is a working example: a starter agent can be cloned and run quickly, while more advanced folders show multi-step reasoning, scheduled execution, or multi-agent collaboration. It covers Claude, Gemini, GPT, DeepSeek, Llama, Qwen, and other open models, with examples using Google ADK, OpenAI Agents SDK, CrewAI, LangGraph, Pydantic AI, Qdrant, and MCP. 6

That makes it more useful than a conventional “awesome” list, but less uniform than a framework. The README says every skill passes a security and evaluation CI gate, while individual templates still have their own dependencies, model choices, and deployment assumptions. The API snapshot shows 1,109 commits, 13 open issues, 124,547 stars, and an Apache-2.0 license. 7 Star it as a pattern library for learning by running; skip treating the catalog's breadth as a common production-quality guarantee for every example.

Vibe-Trading tries to make the trading agent a complete runtime rather than a single strategy script. Its Python 3.11+ backend uses FastAPI, its frontend is React, and the same project exposes a CLI, web UI, REST API, and MCP interface. A connector-first design ties together market data, research goals, backtests, broker access, agent swarms, and execution workflows. The README lists support across equities, crypto, and multiple regional markets, with connectors and data loaders for services such as Yahoo Finance, Futu, Binance, OKX, Alpaca, and others. 8

The differentiation is the boundary between an agent and the financial plumbing around it. A backtesting library owns simulation; a broker bot owns orders. Vibe-Trading tries to own the research-to-execution loop, with session runtime, tool calling, audit records, kill switches, and read-only or paper-only modes for some connectors. That also makes the project a poor candidate for casual live deployment. GitHub lists 797 commits, 67 open issues, 25,288 stars, and an MIT license. 9 Star it for studying connector-first agent architecture; skip handing it live capital until you have independently tested every data, permission, and execution boundary.

DeepTutor treats tutoring as a persistent workspace rather than a chat endpoint. Chat, quiz generation, research, visualization, problem solving, and mastery paths share one agent loop. A local web app and CLI sit on a Python backend with FastAPI and a Next.js 16 / React 19 frontend. The loop can call tools, ground answers in a knowledge base, read attachments, generate images, consult subagents, and write notebook records. Knowledge bases can use LlamaIndex, PageIndex, GraphRAG, LightRAG, LightRAG Server, or Obsidian, while the project describes a three-layer memory system with an auditable evidence trail. 10

The reusable idea is shared learning context. Books, notebooks, question banks, personas, memories, and co-writer drafts can travel between surfaces, so a quiz is not disconnected from the research that produced it. That is a stronger model than a generic RAG tutor, but it also increases operational scope. The API snapshot records 1,075 commits, 56 releases, 55 open issues, 27,928 stars, and an Apache-2.0 license. 11 Authentication is off by default, and Office skills may require subprocess execution in a sandbox. Star it as a serious reference for stateful learning agents; skip it for multi-user deployment until you design the workspace and trust model yourself.

mattpocock/skills packages an opinionated engineering method as small, composable instructions. Its target is the failure before code is written: an agent misunderstands the request, lacks the project's vocabulary, or produces a change that is difficult to test and review. The collection includes workflows such as grill-me, grill-with-docs, test-driven development, bug diagnosis, specification conversion, ticket creation, implementation, and architecture improvement. A setup command asks which issue tracker, labels, and documentation location the project uses, turning generic instructions into local working agreements. 12

Its difference from GSD, BMAD, or Spec-Kit is restraint. The README argues that large process frameworks take control away from the developer and make process bugs hard to repair. These skills can be copied into a project and changed, or installed as a managed Claude Code plugin. The skills.sh installer also supports Codex and other Agent-Skills-standard harnesses, while a native Codex plugin is still on the roadmap. GitHub reports 313 commits, 200 open issues, 177,494 stars, four releases, and an MIT license; the latest release is v1.1.0 from Jul 8. 13 Star it if you want engineering discipline without surrendering control; skip it if you are looking for a turnkey workflow that dictates every step.

Cangjie-skill tackles the gap between consuming knowledge and calling it at the moment of work. It accepts books, long videos, podcasts, interviews, courses, long-form articles, and other source material, then tries to distill reusable methods into an installable skill pack. Its RIA-TV++ pipeline first builds a whole-content overview, runs five parallel extractors for frameworks, principles, cases, counterexamples, and terminology, then applies three filters: independent support in the source, predictive usefulness, and non-obviousness. Approved units are structured as source, interpretation, examples, future triggers, executable steps, and boundaries. 14

That is a different output contract from a summary, a transcript, or a note-taking vault. The repository emits an index, digest, glossary, linked skill files, and test prompts, and it reports a typical three-way screening pass rate of 25% to 50%. The boundary is healthy: material without a repeatable method should not be forced into a skill. At the time of the repository snapshot, GitHub listed 38 commits, seven open issues, 3,828 stars, and an MIT license. 15 Star it if you are building a personal knowledge-to-action pipeline; skip it if your source material is mostly facts or inspiration with no executable method.

OfficeCLI gives agents a document-control surface that standard Office automation rarely provides. It is a cross-platform single binary for Word, Excel, and PowerPoint files, with an embedded .NET runtime and no Office installation requirement. The command model is split into three layers: read the rendered view, inspect and modify a structured DOM, or drop to raw XML. JSON output, batch mode, resident mode, and an MCP server make the same operations available to scripts and agents. Its built-in renderer can emit HTML or PNG so an agent can follow a render, look, fix loop instead of trusting an unrendered document tree. 16

The distinction from python-docx or openpyxl is not just another wrapper. OfficeCLI adds path-based element access, template merge, round-trip dumps, live preview, and a headless workflow for CI or Docker. Compared with Microsoft Office, it avoids COM and desktop installation; compared with LibreOffice, it is shaped around JSON and agent actions rather than a general GUI suite. GitHub lists 5,910 commits, 133 releases, 32 open issues, 19,657 stars, and an Apache-2.0 license. 17 Resident mode can delay writes until a file is saved or closed. Star it for agent-native document manipulation; skip it for unattended publishing until you verify fidelity on your own templates.

UI Skills is a small routing layer for design engineers. Its public interface is intentionally terse: npx ui-skills start routes an agent to the relevant skill set, while categories, list --category motion, and get baseline-ui expose the catalog. The repository is primarily TypeScript and Astro, with the skills stored as a separate directory rather than embedded in a component runtime. That makes it closer to a searchable instruction and reference package than to a UI kit that ships React components. 18

The useful pattern is selection before generation. A design agent can be pointed at motion, baseline UI, or another domain-specific rule set instead of receiving one giant prompt every time. In return, the README leaves more of the actual category content and evaluation method implicit than Hallmark or mattpocock/skills. The API snapshot shows 162 commits, 11 open issues, eight releases, 5,426 stars, and an MIT license; the latest listed release is v0.2.3 from Jun 22. 19 Star it as a lightweight experiment in routing design knowledge to agents; skip it if you need a documented component system, visual regression suite, or production UI primitives.

Codex is the chart's most established terminal agent and the only entry in this week's top ten whose README is explicitly backed by OpenAI. It runs locally, is written mainly in Rust, and is distributed as platform-specific binaries or through npm and Homebrew. The repository includes the CLI, SDK, docs, build tooling, and platform packaging rather than a thin prompt wrapper. Users can sign in with a ChatGPT plan or configure API-key access, then use the agent from a terminal or connect it to an IDE or the Codex desktop experience. 20

Its differentiation is the combination of an open implementation, vendor-backed distribution, and a fast release cadence. The GitHub page lists 8,344 commits, 929 releases, a latest release of 0.144.6 on Jul 18, 99,691 stars, and an Apache-2.0 license. The API snapshot also shows 10,207 open issues, a reminder that scale and visibility create their own maintenance burden. 21 Codex is tied to OpenAI authentication and usage policies, even though the CLI itself is open source. Star it if you want a reference terminal agent and are comfortable with that vendor boundary; skip it if your priority is a model-agnostic or fully offline coding loop.

The strongest pattern is not simply “more AI projects.” It is the conversion of expertise into interfaces an agent can call. Hallmark, mattpocock/skills, cangjie-skill, and ui-skills package design or engineering judgment. DeepTutor and Vibe-Trading wrap long-running domains in stateful runtimes. OfficeCLI and OpenCut make media and documents addressable by tools. Codex supplies the terminal control plane, while Awesome LLM Apps turns many of those patterns into runnable examples.

For a first experiment, start with the entry closest to an existing task: OfficeCLI for document automation, OpenCut for programmable video, DeepTutor for persistent learning, or the skill packs for workflow design. Treat star counts as a discovery signal, then read the README, inspect the license and activity, and test the boundary that matters to your workload before adopting the code.