-
·
Why I Stopped Defaulting to OpenAI
My GitHub history shows an arc from OpenAI to Gemini to Claude — not from conviction but from accumulated experience, with OpenAI still winning specific jobs like Whisper and long-running terminal tasks.
-
·
MCP Is the Glue I Didn’t Know I Needed
MCP matters not as a spec but as the practical layer that stops you from copy-pasting between apps — the moment it becomes real is the first time an agent publishes to Notion, sends a Telegram message, and reads your calendar without you touching any of them.
-
·
What Cowork Changed About How I Run Agents
Cloud-executed agents removed the laptop-must-be-open constraint and changed not just convenience but which tasks I am willing to assign — shifting the relationship from supervision to management.
-
·
Building a Chief of Staff Out of Markdown Files
Two-tier markdown memory beats vector databases for personal and small-team agent memory because it is transparent, free, version-controlled, and never drifts.
-
·
My Claude Code Setup, In Plain English
Six named subagent files, two-tier markdown memory, slash commands, MCP integrations — and every structural choice was made because of something that broke without it.
-
·
Claude Code, Codex, and Cursor: Three Tools, Three Jobs
They are not competing tools. They solve different problems. Knowing which one to reach for depends on whether you are building a system, editing code in an IDE, or running tasks you do not need to supervise.
-
·
The Harness Matters More Than the Model
Eight months of building with Claude Code taught me that the model is maybe 20% of the outcome. The harness is the other 80%.