Writing
Essays and field notes on operating, AI, and building ventures.
Building in the Room
Most consultants observe from the sideline. The best operators sit in the room, feel the tension, and build through it. Here's what that actually looks like.
Read more →Essays
Long-form arguments on operating, AI, and the boundary between them.
MCPs Are a Transitional Technology
MCPs will stick around for OAuth dances and stateful protocols. Everything else is already dead weight. Don't build infrastructure for problems the model has already outgrown.
Read more →Your AI Coding Assistant Forgets Everything. Here's How to Fix That.
I built an open-source logging system for Claude Code that captures every decision as it happens. After 2 months and 38+ sessions across 5 projects, here's how it works and what I learned.
Read more →Systems That Survive You
The best operational systems don't need you to run them. Notes on building infrastructure that outlasts the builder.
Read more →What AI Can't Replace
AI handles the repeatable. Humans handle the ambiguous. The operators who understand this boundary will build the next generation of companies.
Read more →One Operator, Five Projects
An e-commerce brand, a multi-client consulting firm, a partner agency, a consumer brand, and a personal life. Five projects, one workspace, one operator - and about ninety minutes a week of overhead.
Read more →Field Notes
Short, specific observations from building AI workspaces in production.
Don't Tunnel Dev Servers: A 90-Second Near-Miss
I tried to expose a localhost dashboard through a Cloudflare Tunnel. Discovered mid-verification it was actually serving my entire ~/Documents folder to the internet for ninety seconds.
Read more →Convergent Logging: Your AI Workspace Has Two Histories
Interactive sessions log one place, daemons another. The fix isn't smarter logging - it's making both write the same files in the same format.
Read more →Ace: A VS Code Extension for Markdown Annotation and Feedback
Code review for prose. Ace lets you annotate markdown files in VS Code with inline comments, highlights, and threaded feedback - built for AI-assisted writing workflows.
Read about this project →Claude Code from Slack: Why I Built a Local-First AI Bridge
Most AI-in-Slack integrations send your code to a cloud API. I built one that keeps everything local - and it's been running since February.
Read about this project →Local-First Cron Architecture for AI Workspaces
Anything that can use your local tools must run on your machine. Anything reliable must not depend on your machine being on. The fix: GitHub as a task queue.
Read more →What I Check Before Shipping a Workspace
Before handing a workspace to anyone - client, team member, collaborator - I run this checklist. A leaked secret in a markdown file isn't just embarrassing; AI agents treat every file as authoritative context.
Read more →I Built a Pipeline That Couldn't Deploy Itself
When you build a system that auto-deploys updates, the very first deployment can't use that system - the thing doing the deploying is the thing being deployed.
Read more →VS Code's CustomTextEditorProvider Is the Right API for Preview Extensions
When you need a webview that replaces the default editor for a file type, the API exists. Don't hack it with WebviewPanel + closeActiveEditor.
Read more →iCloud Path Resolution Breaks AI Tool State
macOS updates can silently change how ~/Documents resolves. AI tools that key state by filesystem path lose all your project data overnight.
Read more →macOS Python Quirks for AI Workspaces
Traps I hit building the Alcanah daemon and menu bar statusbar. System Python lag, LaunchAgent restrictions, and the NSApp timing trap that ate an evening.
Read more →Git Branches Don't Belong in Brain Repos
If the repo is the thinking, there's no staging for a plan document. Branches solve a review-gate problem brain repos don't have.
Read more →Suggest, Don't Block: AI-Assisted Product Management
Using AI to suggest structured processes without gating work behind them. Process sticks when it saves time.
Read more →A Goal System Your AI Can Actually Read
Most goal systems are for humans to read and AI to ignore. You don't need a separate AI-friendly format. You need a single format explicit enough for both.
Read more →Your AI Tools Are Silently Eating Your CPU
Found 11 orphaned MCP processes burning 550% CPU. The oldest had been running for 48+ hours. Here's the detection logic and the fix.
Read more →Claude Code's Plan Mode Has a Memory Leak
Plan mode saves files with no completion lifecycle. After context compaction, stale plans resurface and Claude starts re-doing work you finished days ago.
Read more →Agent Reliability Patterns
Three techniques for ensuring AI agents complete multi-step workflows without dropping tasks. Defense in depth: awareness, prevention, enforcement.
Read more →