Today's GitHub trending reflects a shift in how the developer community thinks about AI coding agents — the question is no longer just "can an agent do this?" but "will it do it well, cheaply, and with accountability?" Three very different projects point in the same direction: raising the quality floor on AI-assisted work, cutting its cost, and pulling back the curtain on how these systems are actually instructed to behave.
addyosmani/agent-skills: Engineering Discipline for AI Coding Agents
Addy Osmani, longtime Chrome DevRel and web performance advocate, open-sourced a collection of 24 structured engineering workflows for AI coding agents. The skills are organized across six phases — Define, Plan, Build, Verify, Review, and Ship — and encode practices like test-driven development, security hardening, and proper code review that agents tend to skip when left to their own defaults. Rather than a generic prompt collection, this is an attempt to bake senior engineering judgment into the agent layer itself. It picked up 2,656 stars today and sits at 57,428 overall.
github.com/addyosmani/agent-skills
JuliusBrussee/caveman: Slash Your Claude Code Token Costs by ~65%
Caveman is a Claude Code skill built around a deceptively simple idea: instruct the agent to "talk like caveman" — drop filler words, use fragments, stay precise — and you cut output tokens by roughly 65% without meaningful loss in technical accuracy. The project also includes a caveman-compress command that applies the same compression to memory files, shrinking the input token load across sessions. With AI agent costs scaling directly with token volume, a straightforward convention that removes overhead without removing substance is genuinely useful. The repo sits at 72,027 stars.
github.com/JuliusBrussee/caveman
asgeirtj/system_prompts_leaks: A Public Archive of AI System Prompts
This repository has quietly assembled extracted system prompts from Claude, ChatGPT, Gemini, Grok, and other major AI assistants, and it keeps getting updated. At 41,925 stars, it has clearly struck a nerve — there is genuine public appetite to understand how these systems are actually instructed to behave, not just how their makers describe them. The collection is useful for researchers studying AI alignment and behavior, developers debugging unexpected model responses, and anyone curious about the gap between a product's marketing and its actual instructions. It raises real questions about transparency norms in the AI industry.
github.com/asgeirtj/system_prompts_leaks
refactoringhq/tolaria: A Desktop App for Owning Your Knowledge
Tolaria is a cross-platform desktop application for managing markdown-based knowledge repositories — notes, documentation, AI memory files — with full offline functionality and git-backed version control. It gained 369 stars today. The emphasis on user ownership (plain files, no cloud dependency, AGPL-licensed) positions it as a deliberate alternative to notion-style tools that keep your data on someone else's server. As AI agents increasingly write to and read from persistent memory files, having a clean local interface for that layer is becoming more practical than it sounds.
github.com/refactoringhq/tolaria
browser-use/browser-use: Making the Web Navigable for AI Agents
Browser-use continues to trend in the Python ecosystem at 98,569 stars. It gives AI agents a clean interface for navigating and interacting with websites — clicking, filling forms, extracting structured data — without requiring developers to write brittle CSS selectors or custom scrapers. The maintained trajectory of this project alongside the growth of agent skills frameworks suggests that the "agent that can use a computer" pattern is solidifying into a standard building block rather than a research novelty.