Today's most-starred projects share a common concern: improving the scaffolding around AI rather than the models themselves. The biggest mover of the day is a diagram skill for Claude Code, but the broader pattern is about what developers need once a capable model is in the loop — better memory, better tooling, and better ways to package and share reusable instructions.
cathrynlavery/diagram-design: 29 Editorial Diagram Types for Claude Code
This skill ships 29 diagram types as a single Claude Code plugin — everything from architecture diagrams and sequence flows to quadrant plots, Gantt timelines, and radar charts — all rendered as self-contained HTML and SVG with no external dependencies. What sets it apart from the default Mermaid path is deliberate visual design: the output avoids the generic rounded-box aesthetic and can automatically adopt your brand colors and typography after a one-time onboarding step. With over 4,400 stars gained today alone, it's clearly filling a gap developers have been working around manually.
github.com/cathrynlavery/diagram-design
NousResearch/hermes-agent: An Agent That Learns From Its Own Experience
Nous Research released Hermes Agent, a multi-platform AI agent framework designed to run on servers, VPSs, or serverless infrastructure rather than on a local machine. The technically interesting piece is its closed learning loop: the agent autonomously writes and refines its own skills from completed tasks, maintains a searchable memory across sessions, and builds a running model of user preferences over time. Seven execution backends — including SSH, Docker, and Modal — let the agent's environment hibernate and wake on demand, with support for CLI, Telegram, Discord, and Slack out of the box.
github.com/NousResearch/hermes-agent
anthropics/skills: Anthropic's Public Skill Library
Anthropic's official skills repository continues growing as a reference point for anyone building with Claude Code. The repo covers reusable skill packages for document generation (DOCX, PDF, PPTX), creative and design workflows, web testing, and enterprise communications — all structured as self-contained folders with a SKILL.md instruction file. Beyond being a practical library, the repository doubles as documentation for building custom skills and is now the canonical reference for the Agent Skills standard at agentskills.io.
unslothai/unsloth: A Local Studio for Running and Training LLMs
Unsloth offers a desktop application and web interface for running inference, fine-tuning, and deploying models entirely on local hardware — no cloud required. The platform supports a wide range of models including Qwen, Gemma, DeepSeek, and diffusion models across NVIDIA, AMD, Intel, and Apple Silicon GPUs. The notable hook for developers is the unsloth start command, which exposes a locally-running model as an OpenAI-compatible API endpoint that AI coding agents can route requests through — letting teams swap a commercial model for a locally-tuned one without changing any integration code.