Today's GitHub trending tells a consistent story: developers are no longer just experimenting with AI coding agents — they're building the plumbing around them. The projects surfacing this week focus on making agents smarter, more context-aware, and more composable, rather than on the models themselves. The "glue layer" between LLMs and real developer workflows is maturing fast.
cathrynlavery/diagram-design: 44 Diagram Types as Agent Skills
This HTML repo packages 44 diagram types — flowcharts, sequence diagrams, ER diagrams, system maps, and more — as self-contained HTML and SVG files that AI coding agents can consume directly as skills. Rather than asking an agent to figure out diagram syntax from scratch, you hand it a precise template and let it fill in the structure. The result is dramatically more consistent output and less back-and-forth on layout. It's a compact resource that punches well above its file size.
github.com/cathrynlavery/diagram-design
mksglu/context-mode: Smarter Context Windows for AI Coding Agents
One of the more practical problems with long agent sessions is context bloat — tool outputs, intermediary reasoning, and stale file reads pile up and crowd out what actually matters. context-mode is a TypeScript tool that addresses this by sandboxing tool output and persisting session memory across turns, so agents carry forward what's useful and discard the rest. For developers running extended autonomous coding sessions, this kind of infrastructure is increasingly the difference between a session that stays on track and one that drifts.
github.com/mksglu/context-mode
anthropics/knowledge-work-plugins: Open-Source Plugins for Claude Cowork
Anthropic has open-sourced a collection of plugins aimed at knowledge workers using Claude Cowork. The Python-based repo serves as both a reference implementation for plugin authors and a starting set of capabilities for teams that want to extend Cowork without building from scratch. With nearly 29K stars and 625 gained today, it's clearly resonating — the "knowledge worker" use case for AI tooling has lagged behind coding assistants, and a curated, official plugin library helps close that gap.
github.com/anthropics/knowledge-work-plugins
multica-ai/andrej-karpathy-skills: A Single CLAUDE.md to Improve Claude Code
This one is unusual: a single CLAUDE.md file — no code, no dependencies — designed to meaningfully improve Claude Code's behavior based on carefully observed LLM coding pitfalls. Despite its minimal packaging, it has over 218K stars, making it one of the most-starred repositories on GitHub. It's a striking data point about how much developer appetite exists for community-curated prompting patterns. The file distills observations about where code-generation models commonly fail and provides targeted guidance to steer around them.
github.com/multica-ai/andrej-karpathy-skills
hugohe3/ppt-master: Generate Native PowerPoint from Documents
ppt-master takes a document or a topic description and generates a fully-formed PowerPoint deck — complete with charts, speaker narration, and custom templates — using AI. Unlike tools that output markdown-to-slides, it produces native .pptx files with proper formatting, which matters for anyone who actually has to send a deck to a client or present to stakeholders. The Python project has nearly 60K stars and continues to gain traction as AI-powered document generation moves from novelty to workflow staple.