JanusPath

A reconsideration of the organizational paradigm of software development. 1. Why Talk About Organizations Rather Than AI?Over the past two years, most discussions about AI in software development have focused on tools: Can AI write code? Can AI improve development efficiency? Will AI replace programmers? All of these questions matter. But sustained practice has led us to an increasingly strong conclusion: The biggest change AI brings may not be how software is developed, but how developm...

The more I use Codex in real development work, the more convinced I become of one thing: AI is changing how software gets built, but it is not making decades of software engineering thinking obsolete. If anything, the opposite may be true. As agents become better at understanding requirements, modifying code, running tests, and autonomously driving complex tasks forward, many practices originally designed to improve developer productivity and control complexity are beginning to reveal a new k...

Recently, I open-sourced a local-first password and token manager called KeptNear. My initial reason for building it was simple: I wanted a local password manager with a friendly interface—one whose data genuinely remained under the user’s control. Later, as I used Codex to handle more and more real work, I discovered another problem: Agents need credentials too. GitHub Tokens, API Keys, database passwords, cloud platform Access Keys—these credentials are usually scattered across environment ...

On cognitive continuity between humans and AI After using ChatGPT for a long time, I gradually began to experience something I had rarely felt with other tools: it was not merely remembering me. It was beginning to think along my lines. I once gave the same question to ChatGPT and to a GPT model called directly through the API. Both used highly capable models, but their answers were noticeably different. The API response was not poor. It simply resembled a general answer written for anyone....

Over the past year, I have used AI agents such as Codex and Claude Code intensively in my day-to-day development work. At first, like many people, I kept trying to give the agent more capabilities. Whenever I encountered a new kind of task, I added a Skill. Whenever I discovered a rule, I wrote it into AGENTS.md. After a difficult investigation, I wanted to preserve the process so that it could be reused next time. I assumed the agent would become smarter and smarter. Eventually, I discovered...

Much of the time, we believe we are solving a problem. The more accurate description may be that we are working inside a problem that has already been organized for us. Consider a delayed project. On the surface, “someone failed to deliver on time.” We naturally ask: Who was responsible? Who delayed the schedule? How do we prevent it next time? Those questions are meaningful. But they remain at the level of content. Look deeper and the delay may have arisen from unclear responsibility boundar...

Recently, while using Codex to assist with development, I did something that seemed ordinary. I organized the development constraints, design habits, and engineering conventions I use most often into a Skill, allowing Codex to follow them automatically while executing tasks. It began as a simple cleanup exercise. Halfway through, I suddenly realized: Perhaps the most important production activity in future software engineering will no longer be writing code, but extracting rules. The indust...

1. PurposeThis SOP defines how a team should use Codex, OpenSpec, and Superpowers for requirements discovery, product planning, implementation, testing, validation, operations analysis, troubleshooting, and knowledge capture. This SOP is not only for software engineers. It also applies to product managers, QA engineers, designers, operations teams, customer success teams, implementation and delivery teams, data analysts, and any business team that needs to preserve project context. Core goals...

I have seen many teams fall apart not because the business was difficult or people were unwilling to work hard, but because of a persistent, low-intensity sense of suffocation. No KPI can easily measure this feeling, yet it slowly grinds a team hollow. People become quieter, more conservative, and increasingly focused on “just not making a mistake.” Only when a key person resigns does the manager realize that the real cost has been accumulating invisibly for a long time. My clearest understan...

Recently, AI agents have often been described as systems that can “take over everything.” They can write code, run projects, and execute complex workflows automatically. Some people even believe they can already replace engineers. After using them in practice, however, I have reached almost the opposite conclusion: Agents are powerful, but their boundaries are just as clear. 1. A Context Window Is Not MemoryPeople often point out that: “Models now support 200k-token context windows, or ev...