Stavros Ioannidis details his LLM coding workflow
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Stavros Ioannidis published a detailed blog post walking through his LLM-assisted software development workflow, including a real annotated coding session at the end
- Ioannidis reports that around Codex 5.2 and more recently Opus 4.6, he now writes software with a defect rate "significantly lower than if I had hand-written the code," sustaining projects at tens of thousands of lines for weeks
- He built multiple real, maintained projects this way — Stavrobot (a security-focused LLM personal assistant alternative to OpenClaw), Middle (a voice note pendant that transcribes and POSTs to a webhook), Sleight of hand (an art clock with irregular ticks), and Pine Town (a multiplayer drawing canvas)
- His harness of choice is OpenCode, which he requires to support multiple models from different companies and custom agents that can autonomously call each other — features he says first-party CLIs like Claude Code, Codex CLI, and Gemini CLI fail to provide
- His workflow splits work across agents — an architect using Opus, a developer using Sonnet, and one to three reviewers on different models, on the principle that a model "tends to mostly agree with itself" in self-review
- He specifically notes Codex 5.4 is "nitpicky and pedantic" (good for review), Opus 4.6 decisions correlate well with his own, and Gemini 3 Flash has surfaced solutions other models missed
- Ioannidis observes that his engineering role has shifted from writing correct code to architecting systems correctly, and projects in unfamiliar tech stacks (like mobile apps) still become "a mess of bad choices"
Why it matters: Ioannidis lays out a concrete, replicable multi-agent workflow — Opus for architecture, Sonnet for implementation, different models for review — to keep LLM-generated codebases maintainable at tens of thousands of lines. He addresses the common "toy script" critique by listing four shipping, daily-driver projects, and says the engineering bottleneck has shifted from writing correct code to choosing architectures and chaining models well.
