Author: Christoph Nakazawa
\nFirst published: June 3, 2026 · Updated: September 24, 2026
\nSource: cpojer.net
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 I knew it would be outdated quickly, but I didn’t realize it would be outdated that quickly.
\nI actually cannot believe that I rarely write code by hand anymore. Programming has fundamentally changed and I’ve been wondering which engineering values still matter.
\nRecent contributions (mostly AI-written): Vite+ features in Rust; fate 1.0 live views / drizzle / GraphQL; Codiff; Athena Crisis features and 70+ bugfixes; Void metaframework.
\nOverall, 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 all.
\nNow I’m using the Codex CLI with GPT 5.5 high. I prefer the CLI because it starts with a clean slate. Lower reasoning results in poor quality; xhigh is too slow. I’m ineffective at running multiple agent sessions in the same project — I still review all code. I usually have one Ghostty window per project and work on 3–6 projects at a time.
\nWhen fixing bugs, I force a failing test first. Keeping models on a tight leash is still best. I review with Codiff (codiff -w). The best next step is to push to main — PR/collaboration phases often lose the velocity LLMs give you.
Agents amplify ownership. The most effective teams will be small, often two to three people, with clear ownership boundaries and isolated repositories. Strong ownership means understanding architecture, product requirements, tradeoffs, and long-term direction — reviewing agent work and deciding confidently.
\nWith agents, everyone can generate a lot of bullshit. Good taste is about what is worth spending time on.
\nWorking with coding agents is like onboarding new engineers without context constantly. Lint rules, automated testing, and fast verification make agents complete work in 1 minute instead of 60. Tools should operate on changed files.
\nPut relevant context locally in the repo — taste, values, principles — so each agent session (like a new employee) can find it. Simpler, less code is easier to understand and fix.
\nAgents change the cost of owning core libraries. Depending on third-party stacks locks you into decisions you don’t control. Owning your stack gives agents tight constraints and you full product control.
\nAny change should unlock more options for future improvements. Don’t vibe yourself into a corner.
\nEngineering management is going to become more technical, not less. As execution becomes cheaper, managers must retain domain expertise and provide technical leadership — closer to Tech Lead Management.
\nIn the past 30 days: ~770 commits and ~15k lines modified/day on average vs ~328 commits and ~5k lines/day a year earlier — more than twice as many commits and about 3× the code volume vs two years prior, at higher quality.
\nPeople are concerned about jobs. Unless models receive innate motivation aligned with goals and taste, humans stay in the loop — bottlenecked on judgement, not typing. Let’s ship!
","headings":[{"level":1,"text":"Modern Engineering Values","id":"modern-engineering-values"},{"level":2,"text":"How I use LLMs Now","id":"how-i-use-llms-now"},{"level":2,"text":"Modern Engineering Values","id":"modern-engineering-values-2"},{"level":3,"text":"Strong Ownership","id":"strong-ownership"},{"level":3,"text":"Taste, Taste, Taste","id":"taste-taste-taste"},{"level":3,"text":"Strict Guardrails & Fast Feedback Loops","id":"strict-guardrails-fast-feedback-loops"},{"level":3,"text":"Context in the Repo","id":"context-in-the-repo"},{"level":3,"text":"Own your Stack","id":"own-your-stack"},{"level":3,"text":"Option Value","id":"option-value"},{"level":2,"text":"On Management","id":"on-management"},{"level":2,"text":"How much faster?","id":"how-much-faster"}]}}