GitHub Trending Weekly Digest — July 6-11, 2026

The Week in Brief

July 6-11 saw GitHub trending consolidate around a clear theme: AI agent maturity. Not the chatbot kind—the kind that writes code, manages development workflows, and automates away repetitive engineering tasks. The ecosystem is moving fast. Five of this week's top projects are direct infrastructure for AI agents: workflow skills, memory systems, office automation, desktop control, and design-to-code pipelines. The message from developers is unmistakable: we're past experimentation. Now we're building production systems.

The Persistent Champions (Appeared 6+ Days)

Agent Skills 🏆

🔗 github.com/addyosmani/agent-skills

What it does:
A production-grade engineering skill library with 24 structured workflows covering the entire software development lifecycle—from requirements and design through testing, code review, performance optimization, and shipping. Integrates with 70+ AI coding agents (Claude Code, Cursor, Codex, Copilot, Gemini CLI, Antigravity, OpenCode, and more).

Why it matters:
AI agents are great at writing code but terrible at following process. They skip specs, tests, security reviews, and documentation—the things that distinguish production code from prototype code. Agent Skills hardcodes the discipline of senior engineers into reusable workflows. Use /spec to force requirements clarity, /test for TDD rigor, /review for five-axis code review, /ship for safe deployment. This is not a library of functions. It's a behavioral codex.

Tech stack:
Markdown-based skill definitions (SKILL.md format) with Slash command integration. CLI-first, multi-agent aware. Supports automatic skill activation based on development phase. Includes anti-rationalization tables—common shortcuts developers try to take, and why they fail in production.

Strong Runners (Appeared 3 Days)

System Prompts Leaks

🔗 github.com/asgeirtj/system_prompts_leaks

What it does:
Open-source repository documenting the system prompts of 50+ AI products: Claude (Fable 5, Opus 4.8, Sonnet 5, Claude Code, Claude Design), GPT (GPT-5.5 Thinking/Instant, GPT-5.4, Codex), Gemini (3.5 Flash, 3.1 Pro, Gemini Diffusion), Grok, Copilot, Cursor, Meta AI, Perplexity, and more. Includes version diffs tracking how system prompts evolve between releases.

Why it matters:
AI model behavior is shaped by system prompts—the hidden instructions fed to the model before users interact with it. Most companies keep these secret. This project makes them transparent, enabling security research, understanding model constraints, and learning how different vendors position their AI products.

Tech stack:
Pure Markdown documentation with Git version control. Diff utilities for comparing prompts across versions. Web-based diffchecker links. Starred graph visualization.

AI Job Search

🔗 github.com/MadsLorentzen/ai-job-search

What it does:
An AI-driven job application framework built on Claude Code. Automates the entire hiring process: profile creation, job scraping and ranking, CV and cover letter generation, interview preparation, and tracking outcomes. Commands include /setup (import profile), /scrape (find jobs), /apply (end-to-end workflow), /interview (mock interviews), /rank (batch scoring), and /upskill (skill gap analysis).

Why it matters:
Job hunting is repetitive, demoralizing, and time-consuming. This project turns it into a structured workflow that Claude Code can execute at scale. It's less about doing the job for you and more about making the job systematic—remove the friction, collect data, and let AI do the non-creative parts. The Drafter-Reviewer architecture ensures quality, and ATS-aware PDF generation prevents applications from being rejected by parsing errors.

Tech stack:
TypeScript + Claude Code + Python + Bun + LaTeX (for CV and cover letter templates) + Web scrapers for Danish job boards (Jobindex, Jobnet). Extensible to other countries.

OfficeCLI

🔗 github.com/iOfficeAI/OfficeCLI

What it does:
The first Office suite designed natively for AI agents. Single-binary tool that reads and writes Word, Excel, and PowerPoint files without requiring Office installation. Features include 350+ built-in Excel functions (auto-evaluated), native pivot tables, Mermaid diagram support, template merging, and HTML rendering—so AI agents can see what they're actually generating.

Why it matters:
Traditional Office automation libraries (python-docx, openpyxl) force AI agents to work blind. They can't see rendering results, so they generate documents that look broken. OfficeCLI includes a built-in HTML renderer so agents can iterate immediately: build → view → fix. Also eliminates deployment friction (no Office dependency, works in Docker, single binary).

Tech stack:
C# + .NET 10 with integrated HTML rendering engine and OpenXML support. CLI-first with JSON output for AI parsing. Auto-installs as Claude Code/Cursor/VSCode skill.

OfficeCLI (2-day streak)

🔗 github.com/iOfficeAI/OfficeCLI

(Already covered above—appeared days 3-4)

Trending Doubles (Appeared 2 Days)

Taste Skill

🔗 github.com/Leonxlnx/taste-skill

What it does:
A design skill library that prevents AI agents from generating bland, generic, "slop" UI. Provides three tunable parameters: DESIGN_VARIANCE (layout experimentation), MOTION_INTENSITY (animation depth), VISUAL_DENSITY (information density). Includes design reference skills that generate Pinterest-style mood boards alongside code implementation.

Why it matters:
AI-generated UI defaults to the mean: sans-serif, centered columns, muted colors. It's safe and it's forgettable. Taste Skill injects design sensibility—unexpected color combinations, sophisticated typography, meaningful motion. The parameters let agents experiment while maintaining coherence. Works with any frontend framework (React, Vue, Svelte) and integrates with Claude Code.

Tech stack:
Multiple SKILL.md versions (v2 experimental, v1 compatible, GPT variant). Framework-agnostic, installed via npx skills add. Generates design reference images that code-generation skills consume.

Codex Plugin for Claude Code

🔗 github.com/openai/codex-plugin-cc

What it does:
OpenAI plugin enabling Claude Code users to invoke Codex for code review, task delegation, and complex problem-solving from within their workflow. Commands: /codex:review (standard review), /codex:adversarial-review (attack your assumptions), /codex:rescue (delegate complex tasks), /codex:transfer (move conversations).

Why it matters:
Claude Code and Codex each have strengths. This plugin enables multi-agent workflows—use Claude Code for orchestration, delegate to Codex for specialized tasks. Useful when you need a second opinion that thinks differently.

Tech stack:
Node.js 18.18+ with local Codex CLI integration. ChatGPT subscription or OpenAI API key required. Async background execution with progress monitoring.

Firecrawl

🔗 github.com/firecrawl/firecrawl

What it does:
Web scraping and extraction API that converts any website into LLM-ready clean data (Markdown, JSON, screenshots). Features include Search (full-text index), Scrape (single page or multi-page), Map (URL discovery), Agent (autonomous data collection), Interact (click/scroll/fill forms), and specialized format support (PDF, DOCX). 96% web coverage, P95 latency 3.4s.

Why it matters:
Modern AI agents need real-time web data, but traditional scraping is fragile: proxy rotation, rate limiting, JavaScript rendering, anti-bot detection. Firecrawl abstracts all of that. Agents can ask "give me product pricing from this site" and get clean JSON back. MCP support and open-source availability make it a natural fit in agent stacks.

Tech stack:
Multi-language SDKs (Python, JavaScript/Node.js, Java, Elixir, Rust, Go). Open-source AGPL-3.0 with commercial hosting available. MCP protocol support and Agent Skill integrations.

DesktopCommanderMCP

🔗 github.com/wonderwhy-er/DesktopCommanderMCP

What it does:
MCP server giving Claude full desktop control: terminal execution with streaming, file system search, code diffing and surgical text replacements, native Excel/PDF/DOCX support, process management, Docker isolation options, real-time file preview, and audit logging.

Why it matters:
Claude as a web app is limited. Desktop Commander turns Claude into a true development environment—it can run your build, search your codebase, edit files, manage long-running processes, and see results in real-time. Bridge between LLM capability and developer workflow.

Tech stack:
TypeScript + Node.js + MCP protocol. Docker isolation, auto-update mechanism, 6 installation methods. Security-first with symbolic link protection and config validation.

TencentDB Agent Memory

🔗 github.com/TencentCloud/TencentDB-Agent-Memory

What it does:
Local long-term memory system for AI agents using 4-layer hierarchical compression: raw conversation → atomic facts → scene blocks → user personas. Supports hybrid retrieval (BM25 + vector + RRF fusion), Mermaid visualization for short-term compression, and zero external API dependency. Published metrics: <1KB Agent token overhead per 10K conversation tokens, 61.38% reduction in total token usage, 51.52% improvement in task success rate.

Why it matters:
Long conversations explode context windows, degrading agent performance. TencentDB solves this with smart compression—convert verbose logs into compact Mermaid diagrams, then compress further into layered abstractions. Each layer is traceable back to source evidence. OpenClaw integration ready.

Tech stack:
TypeScript + SQLite + sqlite-vec. Hybrid retrieval with BM25 + vector search + RRF ranking. Mermaid for short-term visualization, Markdown for long-term storage. OpenClaw plugin + Hermes gateway support.

Stitch Skills

🔗 github.com/google-labs-code/stitch-skills

What it does:
Agent Skills library for Google Stitch (design-to-code platform). Includes three skill packs: stitch-design (design generation, code-to-design conversion, design system management), stitch-build (Stitch → React/React Native/Remotion/shadcn/ui), and stitch-utilities (design markdown generation, prompt enhancement, multi-page web generation). Compatible with Codex, Antigravity, Gemini CLI, Claude Code, Cursor.

Why it matters:
Design and code are perpetually out of sync. Stitch Skills automates the transformation: take a design, extract component structure, generate React code, apply design system tokens, auto-verify layout. The feedback loop is tight—iterate on design, regenerate code, see changes in real-time.

Tech stack:
Agent Skills standard Markdown + Stitch MCP server. React component generation with automatic validation, React Native export, Remotion video generation from design flows, shadcn/ui integration.

Single-Day Standouts (Appeared 1 Day)

PhotoGIMP

🔗 github.com/Diolinux/PhotoGIMP

What it does:
Community patch for GIMP 3+ that remaps the interface, shortcuts, and workflow to match Photoshop. Tool locations, keyboard shortcuts, menu organization—all Photoshop-familiar. One-click install, preserves existing GIMP config.

Why it matters:
Photoshop users have been blocked from switching to GIMP by friction—different muscle memory, different workflows. PhotoGIMP removes that barrier. It's not GIMP pretending to be Photoshop; it's GIMP with a Photoshop-shaped interface. Massive quality-of-life improvement.

Tech stack:
GIMP 3.0+, CSS theming, configuration file customization. Cross-platform (Linux/Flatpak, Windows, macOS). Photoshop official keybindings as reference.

Claude Video

🔗 github.com/bradautomates/claude-video

What it does:
Gives Claude the ability to watch videos. Downloads videos from YouTube, TikTok, Instagram, Vimeo, Loom (300+ sources via yt-dlp), extracts subtitles and key frames, sends rendered video to Claude with selectable quality levels: transcript-only, efficient (<50 frames, ~9.8K image tokens), balanced (100 frames, ~19.7K tokens, default), token-burner (unlimited scene frames). Smart scene change detection eliminates redundant frames.

Why it matters:
Claude can analyze text transcripts but not video. Many debugging tasks (bug repro videos), creative reviews, and content analysis require seeing what's happening on screen. Claude Video fills that gap. Supports timestamp-precise requests (--start 2:15 --end 2:45).

Tech stack:
yt-dlp (video/subtitle download), ffmpeg (frame extraction), Python (orchestration), Whisper API (speech-to-text when subtitles unavailable, uses Groq for speed). Claude vision for image analysis.

Pocket TTS

🔗 github.com/kyutai-labs/pocket-tts

What it does:
CPU-friendly text-to-speech model (100M parameters) achieving 6x real-time speed on MacBook Air M4. Single pip install activation. Supports 8 languages, voice cloning from custom .wav files, streaming audio output, unlimited text input, WebAssembly community port for browser use.

Why it matters:
TTS usually requires GPU or API calls (privacy + cost risk). Pocket TTS works offline on CPU, making it practical for embedded agents, privacy-conscious deployments, and cost-optimized systems. Pre-trained voices: alba, giovanni, lola, juergen, rafael, estelle, anna, azelma, plus 17 others.

Tech stack:
Python 3.10+ + PyTorch 2.5+ (CPU-only). Hugging Face integration. Stream-based inference. Voice state export for faster inference.

U3-SDK

🔗 github.com/SmartlyDressedGames/U3-SDK

What it does:
Complete open-source code for Unturned, a free open-world zombie survival sandbox game. Includes full engine code, assets, mod SDK, and custom server support. Enables community development, mod creation, and learning from production game code.

Why it matters:
Few commercial games open-source their complete codebase. Unturned's release is a learning resource for Unity developers and game design enthusiasts. Community-driven mod ecosystem.

Tech stack:
C# + Unity. .NET runtime. Multiplayer server architecture.

meshoptimizer

🔗 github.com/zeux/meshoptimizer

What it does:
High-performance 3D mesh optimization library. Vertex cache optimization (adaptive triangle reordering), overdraw reduction, vertex fetch optimization, vertex quantization (10-10-10 SNORM, Half precision), mesh simplification, meshlet clustering for modern Mesh Shaders, and compression encoding. Cross-platform (C++, Rust FFI, JavaScript bindings). Supports GPU-specific optimizations (NVidia Turing+, AMD RDNA2+).

Why it matters:
GPU rendering bottlenecks aren't always compute—they're memory bandwidth and cache efficiency. meshoptimizer provides data-structure optimization before shaders even run. Used in game engines, AR/VR, and Web 3D (glTF).

Tech stack:
C++ library + CMake. Rust FFI (meshopt crate). JavaScript bindings. Command-line tool (gltfpack).

Weekly Themes

🤖 AI Agent Maturity

The #1 theme this week: agents need infrastructure. Not just language models, but workflows (Agent Skills), memory systems (TencentDB), desktop control (DesktopCommanderMCP), office automation (OfficeCLI), and design automation (Stitch Skills). The ecosystem is building scaffolding around agent capability.

🎯 Workflow Engineering

From job search automation to code generation pipelines to design-to-code, the week consolidates around structured workflows. Slash commands, skill definitions, and hierarchical task breakdown are becoming standard. No more "write me a React component"—now it's "execute /spec → /plan → /build → /test → /review → /ship".

📚 Transparency & Introspection

System Prompts Leaks normalizes prompt inspection. Agent Memory makes reasoning traceable. These projects share a philosophy: opaque systems are fragile. Understanding the model's instructions, the agent's reasoning, and decision trails is table-stakes for production.

🎨 Design as Code

Taste Skill, Stitch Skills, and Claude Video suggest a convergence: design and development are merging. AI agents need to understand both visual intent and code implementation. The next wave of productivity gains comes from end-to-end design-to-deployment automation.

🔒 Privacy-First Tools

Pocket TTS and Desktop Commander highlight a shift: developers want local-first, offline-capable tooling. No cloud dependency, no data exfiltration, no surprise bills. Privacy isn't a feature—it's the default expectation.

Takeaway

This week shows GitHub trending reflecting real developer needs, not hype cycles. The projects are boring in the best way—unglamorous infrastructure that solves concrete problems. You don't star Agent Skills because it's cool; you star it because it prevents you from shipping broken code. OfficeCLI isn't innovative; it's just competent office automation without friction. System Prompts Leaks isn't flashy, but understanding your tools deeply is a prerequisite for using them well.

The broader story: AI agents are graduating from chatbots to engineering systems. The code they write needs to be production-grade. The workflows they follow need to be disciplined. The decisions they make need to be traceable. The tools they interact with need to be reliable. This week's top projects address every single one of those requirements.

The trend is unmistakable: if you're building agent-powered developer tools, now is the time. The demand is clear, and the infrastructure is solidifying.

Compiled by Tommy Zhang | July 12, 2026