Today's GitHub trending mixes AI transparency research with new productivity tooling: a growing public archive of system prompts from major AI providers, an autonomous math modeling agent, and a self-hosted AI CRM are all near the top of the charts. The underlying theme is that AI is pushing deeper into specialized professional workflows while the developer community grows more curious — and more organized — about what's actually running inside the models powering those workflows.
asgeirtj/system_prompts_leaks: A Public Archive of AI System Prompts
This JavaScript repository collects and regularly updates extracted system prompts from Anthropic, OpenAI, Google, and xAI. It picked up 217 stars today. For developers who build on top of these platforms, studying how providers structure their own system prompts is a practical window into prompt engineering at scale — and a reminder that even the most polished AI products are steered by explicit written instructions. The archive also puts quiet pressure on providers to be more transparent about the default personas and constraints their models operate under.
github.com/asgeirtj/system_prompts_leaks
melgarafael/DeskcommCRM: Self-Hosted AI Sales Platform
DeskcommCRM is an open-source TypeScript CRM positioning itself as a self-hosted alternative to commercial sales platforms, with native AI agents and WhatsApp integration built in from the start. It gained 504 stars today — second-highest on the trending list. The self-hosted angle is appealing for teams that want AI-assisted sales tooling without routing customer conversations through a third-party SaaS provider. Native WhatsApp support reflects how dominant that channel has become for business communication across Latin America, Europe, and the Middle East.
github.com/melgarafael/DeskcommCRM
jihe520/MathModelAgent: AI Agent for Mathematical Modeling
This Python project deploys an AI agent to handle mathematical modeling tasks end to end, from interpreting the problem to generating a submission-ready paper. It gained 262 stars today. Mathematical modeling competitions like MCM/ICM require translating real-world problems into formal frameworks, then solving and writing them up — a structured workflow that maps well to what current AI agents do well. Whether the output would score in competition is an open question, but as a stress test of how far AI can carry a structured academic task from start to finish, it's a worthwhile benchmark.
github.com/jihe520/MathModelAgent
multimodal-art-projection/YuE: Music Generation with Symbolic Planning
YuE is a Python music generation system that adds a symbolic planning stage before audio synthesis, resulting in more coherent long-form compositions than direct prompt-to-audio models tend to produce. It gained 210 stars today. Its zero-shot cover capability — generating a cover of a song without any fine-tuning on that specific track — serves as a strong benchmark for how thoroughly the model has internalized musical structure rather than memorized specific songs. For developers interested in the frontier of generative audio, the architecture decisions here are worth studying.
github.com/multimodal-art-projection/YuE
max-sixty/worktrunk: Git Worktrees Optimized for Parallel AI Agents
Worktrunk is a Rust CLI that wraps Git's worktree feature with a workflow explicitly designed for running multiple AI coding agents in parallel on the same codebase. It gained 54 stars today — modest by trending standards, but the problem it solves is genuinely new. Standard Git checkouts create constant friction when several agents are working simultaneously; worktrees give each agent an isolated working directory, and Worktrunk streamlines the overhead of managing them. It's a small, focused tool built for a development pattern that barely existed two years ago.