Blog

Thoughts on AI, technology, and innovation

May 7, 2026AI

Script Writer Agent Skills: The Missing Link Between Blog Content and Video Production

No high-star standalone script-writer agent skill exists on GitHub, but YouTube Script Writer's 4-role pipeline and OpenLucid's platform preset matrix offer excellent patterns. We evaluate the capability gap in structured video/audio script creation and a self-build approach.

May 7, 2026AI

Stagehand & BrowserBase: Should AI Agents Adopt the act/extract/observe Pattern?

Stagehand's act/extract/observe primitives offer a compelling pattern for AI browser automation, but BrowserBase cloud is paid. We compare with agent-browser, browser-use, Playwright MCP and evaluate whether to adopt the pattern or self-build.

May 7, 2026AI

Summarizer Agent Skills: Three Missing Capabilities — Map-Reduce, Keyframe Anchors, and Cross-Source Synthesis

We have 4 summarization skills covering mainstream scenarios, but three key capabilities are missing: Map-Reduce long-document chunked summarization, video keyframe anchors, and cross-source synthesis. We evaluate 6 GitHub projects and 6 summarization patterns with prioritized recommendations.

May 7, 2026AI

Trend Monitor Agent Skills: Can We Build Cross-Platform Trend Intelligence on Free APIs?

The most complete trend monitoring skill suite on GitHub, trend-radar, provides a closed-loop monitor→deep-dive→alert→forage→report pipeline, but depends on a paid MCP server. We evaluate the feasibility of self-building trend monitoring on free public APIs.

May 7, 2026AI

Xiaohongshu Automation Skills: High Compliance Risk, Extract Reusable Browser Components Only

The leading XHS automation project MediaCrawler (18k stars) provides multi-platform crawling, but XHS has no public API — all automation is reverse-engineered, creating significant compliance risk. We recommend extracting only reusable components like Playwright anti-detection, not building a standalone XHS skill.

May 6, 2026AI

Deep Dive: AI Agent Sandbox Execution — AgentScope Runtime vs OpenAI Agents SDK

A source-level comparison of how AgentScope Runtime and OpenAI Agents SDK handle sandboxed code execution, from container backends and security hardening to production deployment patterns. Which approach fits your use case?

May 6, 2026AI

Harness Engineering: The Complete Guide to Building Reliable AI Agents in 2026

Harness Engineering is the emerging discipline of designing the scaffolding — context delivery, tool interfaces, verification loops, memory systems, and sandboxes — that surrounds an AI agent and determines whether it succeeds or fails. This comprehensive guide covers the evolution from prompt engineering to harness engineering, core components, key projects, production case studies, and a practical roadmap for adoption.

May 6, 2026AI

Hermes Agent's Recall Mechanism: A Deep Dive into Source Code

A source-code-level dissection of Hermes Agent's three-layer recall architecture — built-in memory, session search, and Hindsight external provider — plus a systematic comparison with seven competing AI agent memory systems.

May 6, 2026AI

MCP Builder Skill Review — The Missing Scaffolding Tool for AI Agent Tool Development

Investigating the MCP Builder ecosystem — from Anthropic's official create-python-server to community scaffolding tools. We examine what MCP Builder actually is, what it does well, what it's missing, and why Hermes might need its own MCP tool factory.

May 6, 2026AI

PowerPoint Skill Review — Teaching AI Agents to Make Slides That Don't Look Like AI Made Them

A critical review of PowerPoint generation tools for AI agents — comparing python-pptx, pptxgenjs, markitdown, and our existing skill. We test slide quality, AI-specific pitfalls, and whether existing tools can produce presentations that don't scream "generated by a bot."

May 6, 2026AI

Ralph Loop Deep Dive — The AI Coding Loop That Won't Stop Until Your PRD Is Done

A comprehensive deep dive into the Ralph Loop ecosystem — from Geoffrey Huntley's original Bash loop to five production implementations. We analyze the core mechanism, compare implementations, and map out how to bring autonomous loop execution into Hermes Agent's existing infrastructure.

May 5, 2026AI

How We Adapted Quality Playbook into Hermes: Extraction, Not Duplication

A step-by-step account of how we adapted the popular quality-playbook-generator skill into Hermes Skill Graph 2.0. Instead of duplicating 479 lines, we extracted universal forensic methodology into two composable skills while discarding what our existing graph already handled.

May 5, 2026AI

Skill Creator Ecosystem: From 180K Installs to Hermes Skill Graph 2.0

A deep comparison of three major skill creation frameworks: Anthropic Skills (182K installs), OpenClaw Skill Creator, and Hermes Skill Graph 2.0. We analyze their design philosophies, strengths, weaknesses, and how Hermes synthesizes the best ideas into a tiered composition model.

May 5, 2026AI

Agent Skills Ecosystem Survey: What 8 Categories Reveal About AI Agent Evolution

A comprehensive survey of 8 agent skill categories — from tmux session control to Playwright testing — revealing which skills are worth adopting, which are redundant, and what the ecosystem's state tells us about AI agent maturity.

May 5, 2026AI

Running NVIDIA Nemotron Nano Omni 30B on a Single RTX 4090

NVIDIA just dropped Nemotron 3 Nano Omni — the first open 30B-class model with native video + audio + image + text understanding. I ran it on a consumer RTX 4090. Here's how it went: the surprising speed, the Chinese capability, the multimodal limitations, and what this means for local AI.

May 4, 2026AI

AI Music Agent Skills: Evaluating the Landscape in 2026

We survey four AI music agent skills — EsshUwU/music-skill (procedural MIDI), Cynaps3 OpenClaw (Suno+Sonauto), vargHQ/skills (multi-modal), and fltman/suno-skill. Most struggle with paid API walls, platform lock-in, or limited capability. Here's what works, what does not, and why the local procedural approach is worth watching.

May 4, 2026AI

Hermes Novel Generator: Multi-Agent Novel Writing as a Skill

A Hermes Agent skill that generates full novels through a pipeline of specialized agents — Orchestrator, Storyteller, Character Agents, Lore Auditor, Prose Stylist. We compare it against NovelClaw (293 stars) and Morpheus (25 stars) and explore its Scene Sandbox model.

May 4, 2026AI

Integrating humanizer into the Agent Skill System — and the Missing Chinese Piece

We integrated humanizer (29 English AI-writing patterns from Wikipedia) into six skills across our blogging and job-hunting pipelines, and built zh-anti-slop v2 — covering 20 Chinese-specific patterns with a trigger/exclusion/risk framework borrowed from patina — all voice-tiered to avoid stripping professional substance.

May 4, 2026AI

psy-core: Tamper-Evident Audit Trails for AI Agent Memory

psy-core wraps your AI agent's memory operations in a tamper-evident audit chain. Every memory write, every skill change gets a receipt. Here's how it works, what it captures, and why it matters for agents that learn across sessions.

May 4, 2026AI

Telepath: A Memory-Aware Visualizer That Already Knows You

Telepath reads your Hermes Agent's persistent memory before rendering any chart or diagram. The more Hermes knows about you, the fewer questions Telepath asks. Built for the Hermes Agent Creative Hackathon with Kimi K2, it turns vague intent into polished visuals.