Ideas, systems & field notes
Exploring Azure, cloud-native systems, and the tools shaping how we build. Independent perspectives, from the architecture to the workbench.
Field Notes D10
An empirical analysis of 280 sequential benchmark runs across 28 GitHub Copilot-hosted models and 10 scenarios — separating the 8–14s first-token prefill floor from true marginal token streaming speed, verbosity bias, reasoning effort overhead, and polyglot coding performance.
Read the articleField Notes D09
A practical comparison of OpenAI Codex, GitHub Copilot, Claude Code, and Chinese AI coding products through token budgets, AI credits, cache behavior, and real agent-session capacity.
Front A08
A field survey of how agent skills and the Model Context Protocol went from experimental add-ons to a critical attack surface — and how scanning, red-teaming, and runtime gateways are being rebuilt to contain it.
Casebook E10
The complete system prompt for GitHub Copilot CLI - a terminal assistant built by GitHub to help with software engineering tasks
Casebook E09
What it takes to make a coding agent reliably produce structured, evidence-backed research at scale — without a human babysitter and without the website ever calling an LLM.
Casebook E06
