Before GenAI
Prior to the release of ChatGPT and Claude (as well as other LLMs) software architecture and engineering was (and still mostly is) modeled and written by hand by people trained to do so. Entire high school and college curriculum has been constructed around developing new talent. Every aspect of software construction is rooted in decades old traditions.
The evolution of GenAI from an entertaining chatbot to a serious tool has been extraordinary, though real productivity has shown to be an elusive outcome. There is still widespread doubt about GenAIs capability in the real world and this is combined with the unethical and destructive business practices by every company building commercial GenAI platforms.
My Professional Interest in GenAI was Immediaate
I instantly saw the potential and iteratively tested the gaps, strengths, and overall effectiveness in using GenAI in software engineering.
Phase I - ChatGPT, the Early Days (2022 to 2023)
In the beginning of ChatGPT we had the ability to produce all kinds of content, including code. But the results were consistently of poor quality and laughably hallucinagenic. The results would invent code out of thin air that looked legitimate, but had no basis in reality. The lawyers using ChatGPT to write their courtroom briefs were the the first victims of these outcomes.
Phase II - Claude AI Web (September 2023 to April 2025)
I continued to work with ChatGPT and Claude AI to test their ability to handle complex architecture and codebases. This gradually improved, but the primary limiting factor was context window and none of the platforms had one large enough to handle serious software engineering. You could build small websites, but anyting marginally complex would turn the LLM into a cyclical refactoring process that could never move past a point of real completion.
Phase III - Caude Code (May 2025 to December 2025)
The same efforts continued with a CLI tool, but the results did not move the needle in a meaningful way. Context windows were now known to be the primary limiting factor along with the evolution if the models themselves.
Up to this point, I'd spent nearly two years trying to extract reproducible productivity from any LLM I could access and at every point I stopped because they simply could not prove useful.
Phase IV - The Tipping Point - Opus 4.6 and Claude PRO MAX Subscriptions (December 2025)
The world literally turns updside down. Anthropic's release of Opus 4.6 and its subscriptions opening larger context windows removed the complexity contraints and opened up the capability of designing and building complex software with GenAI.
DevArch's Timeline
One of the first repeated tasks I performed even in the earliest days of using ChatGPT was to complete every session with one prompt: "Write a work summary"
This was necessary (and still is) because an LLM has no memory. In order to maintain context over sessions, you need to ask the LLM to summarize everything discussed, designed, implemented, and completed. And my template for work summaries evolved so that reading the previous work summary was the catalyst for the start of every session. I have thousands of these summaries in markdown files.
That corpus of work summaries along with traditional documents like architecture decision records, proposals, plans, acceptance test criteria, eventually led to the idea that my GenAI usage patterns were something I could document and organize.
In February of 2026 I reviewed Claude Code's agent, hook, and skill layout and developed a plugin that leverages those capabilities that became DevArch AI.
Summary
DevArch AI is proven as an effective and productive engineering harness for Claude Code. Although it is opinionated based on my experience as an architect and engineer, it's based on well-known practices. Domain-Driven Design, Test-Driven Development, Method/1, and other product and project standards from many decades are all elements of DevArch.