Blog

Thoughts on AI, technology, and innovation

August 20, 2026Tech

Task Vault: I Built an Obsidian Plugin That Treats Tasks as Files

Changing a task's identity from a line of text to UUID + frontmatter + append-only execution log — and making AI agents first-class task owners. The data model behind Task Vault and 4 non-obvious engineering decisions.

August 14, 2026Tech

Three FFmpeg Audio Mixing Traps: LUFS Parameterization, asplit Ducking, and amix Silent Attenuation from the OpenMontage Source

FFmpeg's filter_complex looks like a pipeline, but a label can only be consumed once. loudnorm looks like it has a sensible default, but -16 and -14 are two different distribution strategies. amix looks like it's mixing, but normalize=1 silently attenuates your narration by -6dB. This post dissects all three traps from OpenMontage's recent fixes and explains how a production-grade AI video pipeline should handle them.

August 14, 2026Tech

FFmpeg Chromakey Green Screen Compositing: The 1×1 Pixel Bug and Alpha Portability Fix

OpenMontage's chromakey processor was silently collapsing every frame to a single pixel. The root cause: FFmpeg's overlay taking the size of its first input, combined with a scale=iw:ih no-op. Here's the full breakdown of both bugs and why yuva420p is the key to cross-build portability.

August 13, 2026Tech

ntfy vs Gotify vs Nostr: A Deep Comparison of Self-Hosted Push — How to Get Your AI Agent's Cron Results Onto Your Phone

You've self-hosted an AI agent — now how do its cron results reach your phone in real time? This deep comparison breaks down ntfy, Gotify, and Nostr Relay across architecture, client coverage, security model, latency, attachments, and message history. The verdict is unambiguous: ntfy for AI agent push, Gotify for Android-only setups, Nostr for the decentralization faithful.

August 12, 2026Tech

Reading bitchat's Source: What's Actually Inside the 35.2k-Star Bluetooth Mesh Chat

bitchat's whitepaper calls its relay a 'controlled flood', but the thing that actually saves airtime is a jitter window that widens with node degree. Its sync filter is lifted straight from Bitcoin's BIP-158. And its identity model is split down the middle: unlinkable keys per geohash on the Nostr side, a never-rotating 8-byte fingerprint broadcast on the BLE side. This is a source-level read of all three, plus where it stands against Briar, Meshtastic, and Bridgefy.

August 12, 2026Tech

A 7,710-Star Diagram Skill With No Renderer: What diagram-design Validates, and What It Doesn't

diagram-design uses 27 Markdown specs to make a model hand-write SVG coordinates. It turns palette, fonts, accessibility and supply-chain rules into hard CI gates — yet ships zero validation for the five connector rules it calls non-negotiable. That gap is the real lesson about agent skills.

August 12, 2026Tech

Inside Orca: When Agents Write the Code, Discipline Has to Become a Machine Gate

stablyai/orca hit 43,205 stars and 8,490 commits in five months, with most of the code written by a fleet of agents. After reading the source, the part worth stealing isn't the features — it's how the team compiled engineering discipline into machine gates: a line-count ratchet that can only shrink, 74 reliability gates that need 100 soak runs to be promoted, and a seven-step protocol just to rename one socket file, because the old shape produced twenty-three defects across seven review rounds.

June 16, 2026Tech

Iroh 1.0: Dial Keys, Not IPs — A Deep Dive Into the P2P Networking Stack

On June 15, 2026, n0 shipped iroh 1.0 — the first stable release of a Rust networking stack that lets you dial a device by its public key instead of its IP address. This deep dive unpacks how iroh hole-punches through NATs, falls back to relays, builds on its own QUIC implementation, and turns the internet into a secure localhost. We cover the architecture, the composable protocol ecosystem, the new official Python/Node/Swift/Kotlin bindings, and what the 1.0 wire-stability guarantee actually means for production.

May 14, 2026Tech

Harness Engineering: A Comprehensive Guide to the Backbone of Modern Industry

A deep dive into harness engineering — the discipline behind every vehicle, aircraft, and industrial machine's nervous system. Covers major industry players from Yazaki to TE Connectivity, key standards like IPC/WHMA-A-620 and USCAR, modern design tools, and the automation trends reshaping the field.

May 6, 2026Tech

Chesterton's Fence: Why You Should Understand Before You Tear Down

Chesterton's Fence is a mental model that advises against removing something until you understand why it exists. Originating from G.K. Chesterton's 1929 parable, it has become a cornerstone of second-order thinking in software engineering, organizational design, and public policy.

May 6, 2026Tech

Code Simplifier Skill Review — Cleaning Up Code Without Breaking Things

A multi-angle review of the code-simplifier skill ecosystem — comparing the Anthropic Claude Code plugin with our adapted Hermes version. We examine the five principles of safe simplification, test both implementations against real code patterns, and identify what's worth borrowing versus what's already covered.

May 6, 2026Tech

Webapp Testing Skill Review — Browser Automation for AI Agents That Actually Works

Hands-on review of the webapp-testing skill — comparing our Playwright-based approach against alternatives in the AI agent ecosystem. We explore the reconnaissance-then-action pattern, server lifecycle management, and why most AI testing tools overcomplicate what should be simple.

May 5, 2026Tech

GitNexus Deep Dive: The 35K-Star Code Intelligence Engine for AI Agents

A comprehensive analysis of GitNexus — the zero-server code intelligence engine with 35K GitHub stars. We examine its 12-phase ingestion pipeline, 16 MCP tools, process detection engine, LadybugDB graph storage, and evaluate its fit for our AI agent skill ecosystem.

May 5, 2026Tech

Tmux Skills for AI Agents: Why We Don't Need a Tmux Wrapper

The tmux control skill has 2.8K installs on skills.sh, but does your AI agent actually need it? We examine the architectural trade-offs between tmux-based session control and native terminal tools.

May 4, 2026Tech

From Sync to Backup: The Missing Half of AI Agent Data Safety

Sync keeps your AI agent consistent across machines, but it cannot save you from a crashed server or a mistaken deletion. A deep dive into why backup is the natural next step after sync, and how to close the gap with a daily ZIP + git push strategy.

May 4, 2026Tech

Integrating Skywork Office Skills into Hermes Skill Graph 2.0

A deep dive into evaluating and integrating SkyworkAI's open-source Office Skills into the Hermes Skill Graph 2.0 system — security audit, multi-dimensional assessment, architecture decisions, and the full adaptation process.

May 3, 2026Tech

Skill Graphs 2.0: A Tiered Architecture for AI Agent Skill Management

Skill Graphs 2.0 is a tiered architecture that organizes AI agent skills into atoms, molecules, and compounds — transforming a flat toolbox into a composable, versioned knowledge graph. We share the methodology, metrics from a 194-skill deployment, and the design lessons learned.

May 3, 2026Tech

Spark 2.0: World Labs Unleashes Streamable 3D Gaussian Splatting on the Web

Spark 2.0 brings Level-of-Detail streaming and virtual paging to 3D Gaussian Splatting, enabling scenes with 100M+ splats to render smoothly on any device. Built by Fei-Fei Li's World Labs, it redefines how 3DGS content is consumed on the open web.

May 2, 2026Tech

The AI Agent Skills Multi-Machine Sync Problem: A Deep-Dive Across 8 Ecosystems and 6 Strategies

When you run the same AI coding agent across MacBook, Linux workstation, and cloud VM, how do skills stay consistent? A systematic comparison of sync strategies across Hermes, Claude Code, Cursor, Copilot, OpenCode, Aider, and Windsurf — plus six general approaches from dotfiles to Nix.

May 2, 2026Tech

How I Installed 10 Hermes Agent Skills from an Awesome List (and Fixed Them All)

I scanned 0xNyk's awesome-hermes-agent list (2,197 stars), identified 11 installable skills across 7 repositories, batch-installed 10 of them, diagnosed every broken dependency, and patched platform compatibility issues. Here's the full playbook.