Walkthrough of the asgeirtj/system_prompts_leaks repo (58K+ stars, CC0). Documents hidden system prompts from every major AI: Claude, ChatGPT, Gemini, Grok, Copilot, Perplexity, Cursor, and more. Reveals how each company configures guardrails, personalities (GPT-5.1 has 8 modes), and tool access (Claude Design: 48 tools, 16 skills). DiffChecker links track guardrail evolution between versions.
Matt Pocock shows how Claude Code's default system prompt ships at ~25K tokens of mostly unused tool definitions. By disabling unused features in settings.json (plan mode, crons, bundled skills, workflows, remote control, AI connectors, artifacts), he cut it to 8K tokens — a 68% reduction that improves output quality.
XBOW topped HackerOne's US leaderboard at $1B valuation — then free open-source tools (Shannon, Strix, Tempest) matched it on its own benchmark. MCP + HexStrike plugs 150+ hacking tools into any AI. The price of autonomous hacking fell from $6,000 to $0.
Chris Ashby walks through creating a professional design system from an image reference using Builder OS and Claude Fable 5 — producing a dual-file spec (design.md + design.html) with claude.md rules to prevent design drift across AI-assisted development.
Complete 103-minute course on Vercel Eve — the Next.js for agents. File-system-first framework where your entire agent is a folder of markdown + TypeScript. Covers: setup, tools (Zod + defineTool), instructions vs skills (on-demand loading), channels (Slack/Discord/Telegram/WhatsApp), MCP connections, sandboxing, durable sessions, sub-agent delegation patterns (fan-out, specialist handoff, ReAct), schedules, and one-command deploy.
Two students open-sourced Strix — an AI pentesting agent that rivals the $270M XBOW startup. With 26k GitHub stars and a real CVE in etcd (CVSS 8.8), it brings autonomous vulnerability discovery to anyone with an API key.
Baidu's Unlimited OCR model reads entire books in a single forward pass using a sliding-window decoder inspired by human 'soft forgetting' — achieving 93.23% SOTA on OmniDocBench V1.5 with only 500M active parameters at inference, and released under MIT license.
Sam Witteveen explores Ornith 1.0, a family of open-weight self-scaffolding LLMs from Deep Reinforce that learn to write their own agent harnesses for coding tasks — featuring four models (9B to 397B MoE), a novel two-stage GRPO training loop, and impressive benchmarks against Claude Opus.
Hadley Wickham demonstrates a three-file workflow for AI-assisted data cleaning — keeping a cleaning script, YAML data dictionary, and Parquet output in sync via Claude Code and MCP REPL, with a live walkthrough using the NYC Elevators dataset.
Nate B Jones introduces Open Skills — a public library of portable, composable agent procedures that solve the procedural debt problem across Codex, Claude Code, and Cursor. Beyond prompts: skills with trigger rules, verification standards, and scope that travel with you.
AI Search reviews GLM 5.2 from Z.AI — the new #1 open-source model that beats GPT 5.5 and Gemini 3.1 Pro on multiple benchmarks. Packed with hands-on demos: interactive 3D Earth, promo videos, ray tracing from scratch, 3D engines, music composition, and Manim animations. Plus how to use it in Claude Code, ZCode, or any agentic harness.
Brainqub3 walks through a real client audit of a 100K-line vibe-coded prototype, proving why line-by-line code review is mathematically impossible (Rice's theorem) and presenting four principles for disciplined architectural review — plus his open-source architecture-as-is Claude Code skill that automates the process.
Steve from Builder.io introduces /visual-plan and /visual-recap — open-source Claude Code skills that replace static markdown plans with interactive MDX documents featuring wireframes, API specs, schema diagrams, and annotated code, arguing engineering is entering a new plan-level abstraction phase.
Cole Medin introduces Omnigent, an open-source meta-harness from Databricks that orchestrates Claude Code, Codex, Pi, and local models under one unified layer — with guardrails, cross-device sessions, and custom orchestrator support.
A comprehensive 14-step roadmap from Codez (@0xCodez) for transitioning from manually prompting AI coding agents to designing automated loops — covering the 4-condition test, five building blocks (automations, worktrees, skills, MCP connectors, sub-agents), minimum viable loops, and critical failure modes like the Ralph Wiggum loop and comprehension debt.
Apple’s WWDC26 session shows how to build fully local agentic AI workflows on the Mac using the four-layer MLX stack. M5 Neural Accelerators deliver 4× faster prompt processing, continuous batching serves parallel subagents, and distributed inference spreads models across multiple Macs via Thunderbolt RDMA. Live demos include building a SwiftUI app from scratch and fixing bugs in Xcode — all locally.
Apple introduces Container Machine at WWDC26 — a new feature combining container speed with VM persistence for seamless Linux development on macOS. Automatic user mapping, shared filesystem, and native integration mean you can edit in Xcode, build in Linux, and test from Safari without copying files or context switching.
Marlene Mhangami and Liam Hampton from Microsoft/GitHub introduce MCP Apps — a new capability that lets MCP server tools return rich, interactive HTML components directly inside the VS Code chat window. Covers the architecture, real-world adopters (Shopify, Excalidraw, Figma), and a live demo profiling a Go app with an interactive flame graph.
Nick Nisi from WorkOS shares a counterintuitive discovery: deleting 95% of his agent skills actually improved performance. He walks through building, managing, and ruthlessly pruning skills across 20+ repos in 8 languages.
Mansel Scheffel breaks down Claude Code's new Dynamic Workflows — a feature that moves sub-agent orchestration from the LLM context window into deterministic JavaScript scripts, giving developers explicit control over state, loops, and parallel execution.
DevOps Toolbox reviews Babysitter, an open-source agent orchestration framework that eliminates hallucinations through deterministic quality gates, enforced TDD, and systematic task decomposition instead of relying on rules.
Matt Pocock examines Cursor's thermonuclear code review skill — an aggressively strict automated review prompt that pushes AI agents to suggest structural rewrites, not just surface fixes.
