Modern Engineering Values | Christoph Nakazawa

‹ homeTable of Contents<br>› How I use LLMs Now<br>› Modern Engineering Values› Strong Ownership<br>› Taste, Taste, Taste<br>› Strict Guardrails & Fast Feedback Loops<br>› Context in the Repo<br>› Own your Stack<br>› Option Value

› On Management<br>› How much faster?

Modern Engineering Values

Published on June 3, 2026 14 minutes reading time, 2700 words

At the end of last year I shared my LLM workflow in You are absolutely right!?. I knew it would be outdated quickly, but I didn’t realize it would be outdated that quickly.

I actually cannot believe that I rarely write code by hand anymore. Or rather, I cannot believe that I used to write code by hand! Programming has fundamentally changed and I’ve been wondering which engineering values still matter.

Before I get started, and to avoid criticism that I haven’t shipped anything, here are my contributions over the past several months:

Vite+: Built features in Rust for the launch (I don’t even know Rust), 90% AI-written

fate 1.0: Added live views, drizzle & GraphQL support, a Vite plugin, and garbage collection, 100% AI-written

Codiff: A fast, minimal, and beautiful diff review app, 100% AI-written

Athena Crisis: New features and 70+ bugfixes, 100% AI-written

Void: Metaframework and Cloud platform (Not yet shipped), 100% AI-written

Overall, I’ve found that coding agents now write code that is as good as or better than what I would write, and they do it in minutes instead of weeks. Since coding is no longer the bottleneck, I get a lot of work done that otherwise wouldn’t have happened at all1I even get to do silly things, like “VoidZero plays Pokémon”, a custom ROM hack that worked over websockets, in about five prompts, for our internal All Hands.

Most of my work is public and you can validate my work directly. I’m not bragging, I’m simply astonished. There has never been a developer experience improvement of this magnitude that allowed me to ship higher quality software so much faster.

Take Athena Crisis for example: It’s a pre-AI codebase that I wrote by hand. I haven’t had time to work on it lately so I had Codex run in a loop to identify bugs and fix them. It fixed 70 bugs, implemented new features, and wrote tests – and I was able to do all that while I was waiting on agents in other projects! As a result, Athena Crisis is now faster, more stable, and has more features than ever before. On top of all that it streamlined CI steps, improved the Steam app upload pipeline, wrote tests for some complex multi-user interactions, and I had it make a custom app to compare 200 pixel art icons with new variants to give the game a new look. None of this would have happened without coding agents.

How I use LLMs Now

Last year I didn’t run models locally, mostly isolated context to a specific module or function, and generated multiple versions for each solution.

Now, I’m using the Codex CLI, with GPT 5.5 high.2I prefer the CLI because it starts with a clean slate. There is no sidebar with old chats I don’t need, and I prefer one window per project instead of all projects in one window. It consistently one-shots correct solutions with very little prompting or structure, as long as the guardrails in the codebase are strong. I found that lower reasoning results in poor quality, and xhigh is too slow and overcomplicates things.

I’m not using /fast because my understanding is that it’s quantized and produces worse output. I tried various other models, but nothing is as fast, as correct, or as easy to work with as Codex.3I appreciate everyone working on local models, and have high hopes for all of them to converge eventually!

I am ineffective at running multiple agent sessions in the same project (with worktrees, copies, or other solutions). I still review all code, and juggling orthogonal code changes in the same project slows me down. Context switching allows me to scale myself better and I can effectively work on multiple projects simultaneously. I usually have one Ghostty window per project, with a Codex tab on the left and a terminal on the right. Now I can work on 3-6 projects at a time, and the bottleneck is usually discussing and reviewing code.

All this multitasking added a spatial dimension to my work. I used to love just working from anywhere using my laptop, but now I prefer having a large screen available and assigning parts of the screen to various projects. For example, fate always ends up in the top center of the screen, void on the left side, and Athena Crisis on the right. Somehow the spatial assignment helps me multitask more effectively.

In terms of prompts, I’m pretty lazy. I often start new sessions for new topics (although context management matters much less with Codex these days) and ask “I want to build X. Get context from Y and Z, make a proposal, and ask me any questions you might have”. I iterate on the proposal by talking to the agent before asking it to execute on the plan.

When fixing bugs, I point...