Field notes · 10 entries
Writing about building.
Practical notes on AI, data, and the tools that make ambitious work feel lighter.
I benchmarked all 17 GPT-6 settings; this is how I use them now
GPT-6 results suggest increasing reasoning effort pays off—but cost and latency rise sharply, and the best model family depends on your budget.
Does your agent need Context7?
A 600-task paired comparison shows how Context7 influences agent behavior, cost and time.
Does RTK really reduce token costs the advertised 60-90% percent?
A 600-task paired comparison shows how RTK changes tool output, message count, and the cost of running an LLM agent.
How I use OpenAI's GPT-5.6, based on 175,000+ tests
Surprising results from looking at the cost, intelligence, and latency tradeoffs between GPT-5.6 capability tiers and reasoning effort.
Stop using MCP, use CLI instead
MCPs are powerful, but their context bloat makes AI agents less efficient. Here’s why CLIs are a better interface for tool use.
Snowball: Self-Improving AI
AI automates a lot, but it quickly devolves into tedious context engineering. Snowball automates the context management engineering.
Slash Commands ate My Workflow
Opencode commands turn my parallel agentic worktree coding sessions into one-liners. 3-4X faster.
Managing Prompts in a Git Repo
How I manage all my prompts in a Git repository.
Parallelize Coding Agents with Git Worktrees
How I get 2-3X the productivity from parallel coding agents.
Welcome to My Blog
Introducing my new blog where I share thoughts on AI, engineering, data science, technology, and more.