A kanban an AI can read and execute

Vibe-coding AIs like Claude Code write code fast, but on a growing project they lose the thread: what's done? what's missing? Ordinary project-management tools don't help, because they're built for a person opening them in a browser — not for a model that has to read and update them by itself while it programs. This guide was born building Ykan, a kanban designed to be read by an AI: here are the conventions that make it possible, applicable to any board.

The kanban as shared memory

The core idea is simple: the board becomes the coordination point between whoever directs the project (you) and the AI executing it. It's not a to-do list for you: it's a memory the AI reads at the start of every session — because a model always starts «cold», with no memory of the previous session — and updates as it works.

The conventions that make a task readable by an AI

A task written for a human and one written for an AI look the same, but the second needs more discipline:

• A title that states the action, not the topic. «Add validation to the login form», not «Login form».
• Enough context in the description. The AI doesn't remember what you said yesterday: each task must contain what's needed to execute it on its own — files involved, constraints, decisions already made.
• Explicit completion criteria. How do you know it's done? If it isn't written, the AI (like a person) guesses.
• Associated files. Naming which files to touch guides the model and reduces errors.

The golden rule: write every task as if whoever executes it knows nothing about the rest of the project. Because often that's exactly the case.

The flow: from idea to code

The typical flow is this. You start from an idea, describe it to the AI and ask it to turn it into tasks. On a large project that can be hundreds of tasks generated in one go — all visible, editable and rewritable by hand before a single line of code is written. That's the moment the human reviews the plan: here you correct, cut, reorder. Then you tell the AI to start: it opens tasks one at a time, executes them and, when done, moves them itself into «Done» or «Review». You stay free to approve or send back any task, without rewriting the plan from scratch.

An interface a machine can use

For the AI to talk to the board without clicking, you need an API: a few simple endpoints — read the board summary, complete a task, move it between columns. The model doesn't «look» at the graphical interface the way you would; it asks the board via API and gets text back. The same logic, exposed as an MCP server, also lets you manage everything from mobile.

Don't trust the move: verify

A common mistake is treating a task as «done» just because it landed in the Done column. An AI (like a hurried colleague) can move it without actually completing it. The countermeasure is verification: having the AI read the files associated with the task and check whether the work was really done, instead of trusting the column. It's the difference between «says it's finished» and «is finished».

Conventions matter more than the tool

You can use Ykan, another kanban or a plain text file: what makes the system work isn't the software, but the formatting conventions — self-contained tasks, clear criteria, an interface the machine can read. Learn those, and any board becomes a tool an AI can genuinely use.