The theme today is craft: the developer community is shipping structured tooling not just to help AI agents do tasks, but to help them do tasks well. Three GitHub projects gaining rapid traction this May Day share a common thread — they codify institutional knowledge (engineering process, financial analysis, design principles) into a form that AI agents can actually use reliably.
mattpocock/skills: Engineering Discipline as Installable Agent Skills
Matt Pocock — best known for his TypeScript education work — has published a collection of Claude Code skills (also usable with other AI coding tools) that target four failure modes he identified in AI-assisted development: misalignment, verbosity, code quality drift, and architectural degradation. Skills like /grill-me have the agent ask probing clarifying questions before writing a single line of code; /tdd enforces red-green-refactor cycles with explicit failing tests before implementation; and /improve-codebase-architecture guides agents through deliberate structural improvements rather than patchwork edits. The package installs in one command via npx skills@latest and picked up over 6,000 stars today — a strong signal that developers are actively looking for ways to make their AI collaborators more methodical and less prone to confident-but-wrong shortcuts.
TauricResearch/TradingAgents: A Trading Firm as a Multi-Agent System
TradingAgents models a financial trading firm as a team of specialized LLM agents: fundamental, sentiment, news, and technical analysts each produce independent reports; a pair of bullish and bearish researchers then debate those findings; a trader agent synthesizes the debate into a position; and a risk management layer reviews every transaction before execution. Built on LangGraph with support for OpenAI, Anthropic, Google, and DeepSeek models, the framework features structured multi-round debate, persistent decision logging that builds on prior trades, and checkpoint-resume for long-running sessions. The project added over 2,000 stars today, reflecting genuine interest in moving multi-agent collaboration beyond chat pipelines and into structured simulations with explicit roles, accountability, and debate at each decision layer.
github.com/TauricResearch/TradingAgents
pbakaus/impeccable: A Design Vocabulary for AI Harnesses
Paul Bakaus built Impeccable to solve a specific frustration: without explicit design guidance, AI coding tools default to the same generic patterns absorbed during training — purple gradients, nested cards, dense over-engineered layouts. Impeccable provides one skill and 23 specialized commands (/audit, /polish, /animate, /critique, and more), each backed by curated anti-patterns spanning typography, color, spacing, motion, and UX writing. It works across Claude Code, Cursor, and GitHub Copilot. The goal is less about automating design decisions and more about giving agents a precise professional vocabulary so developers can redirect them clearly when the output starts looking like every other AI-generated interface.