Here’s a brief note on development — two interesting ideas
Recently, I finally decided to structure my project documentation exactly like OpenAI does in its Engineering Harness — I created a tree of MD files in the /docs folder alongside the code, and scattered AGENTS.MD across folders within the project.
This method works great, especially when using RFCs for planning. Thanks to it, agents can easily find the necessary documentation, quickly solve tasks, and if left unmanaged, can accumulate technical debt.
But this structure has two hidden advantages. Let me explain with examples.
Feature Migration. Currently, I have two key projects — the BitGN platform for agent competitions and a new version of Abdullin Labs, where I plan to publish an English version of the course. They develop at different speeds, use different tech stacks, but what unites them is that both are tailored to my workflow with Codex and Claude.
Sometimes, an opportunity or idea appears in one project that I want to quickly implement in another. For example, on the BitGN platform, I implemented a zero-downtime update mode. It arose because during active operation, the platform sometimes experienced brief downtime. So, I assembled a mechanism based on SystemD Socket Activation:
1. smoothly shuts down the old version of the application (allowing current logic to finish)
2. continues accepting new connections at the OS level
3. creates backups, launches the new service version, and updates the database if necessary
4. connects the application to the same socket, which the operating system continues to hold
Everything works perfectly: I can deploy up to twenty times a day (usually no more than ten), at any time, and no one notices.
Of course, such a feature is quite complex: it requires changes in the application, server configuration (registering the socket in SystemD and connecting the service), and a strict sequence of commands during deployment.
When I wanted to quickly transfer this feature to Abdullin Labs, I did the following: documented everything in /docs — first described the main principles at all levels, then added details specific to the project. Then I took the new feature from another project and adapted it to current conditions — so I ended up with updated configs and scripts for make install && make deploy. It automatically deploys the new server configuration and fresh application version.
This idea grew out of another important point — rapid startup of new projects: now I want to try launching my MCP server, which will allow agents to work with shared memory (shared memory bank), control changes via event sourcing, and roll them back if needed. I also plan to embed optimized context graphs from my reasoning project.
To avoid building everything manually from scratch each time, I asked one of my projects to prepare an RFC — a document that will help agents launch such a project from scratch. In another project, I simply created an empty folder and asked Codex to deploy it. Then I went inside the folder, activated the Nix environment — and got a complete stack with all settings tailored to my preferences (even took screenshots). Everything works automatically and conveniently.
If later I want to add new features — I can easily drag and drop them using the “Feature Drag-and-Drop” approach.
It’s amusing to observe that earlier, during the launch of a Data Science department for a large international logistics company, we spent a lot of time creating complex cookiecutter templates and automated processes — but now it’s enough just to ask Codex.
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby