Preparing a project for AI agents shouldn't be a manual ritual
Every time I start a new project, the same thing happens. Before writing the first line of real code, I have to sit down and explain to whichever AI agent is on duty how everything is put together: what conventions the repo uses, where everything lives, which rules aren't negotiable. It's a ritual that repeats project after project, and every time I do it a little differently, a little by hand, a little from memory.
That made me think the problem wasn't just mine. Setting up CI/CD for a new project stopped being optional a long time ago. Leaving a project ready for an AI agent to understand should have the same status — not an improvisation every time, but a setup step with the same seriousness as any other.
With that idea I built a framework meant to integrate into any
development project, regardless of stack. Installation can be manual,
via automated scripts (.bat for Windows, .sh for Linux and Mac), or
with a single curl command for the quick version. I documented support
for 23 different AI agents — Copilot, Claude, ChatGPT, Cursor, among
others —, more than 15 databases, and templates for more than 7 MCP
configurations.
Note
The goal isn't to "talk nicely to the AI." It's to leave a project in a state where any agent — today or a year from now, with a different model — can understand it without someone having to sit down and explain everything all over again.
To show it in action I built a concrete example: BackupGuard Pro, a fictional backup monitoring system set up with Docker Compose, designed specifically to give the framework something real to stand on — not an empty "hello world," but a project with pieces that interact with each other.
The project set out to document 87 technologies with verified official sources. It's not a number I can reconfirm today with the same accuracy I wrote it with at the time — but I can say what it was meant for: every documented technology had to have a real reference behind it, not a model hallucination about how some library "probably" works.
Leave a project without CI/CD and you accept something is going to break in production sooner or later. Leave a project unprepared for AI agents and you accept the same thing, just at a different layer: every new agent will improvise its own understanding of the code, with the same variability — and the same risk — as any configuration that was never documented.
I originally published this on LinkedIn