AI Context Kit: One Git Repository with Markdown Instead of Copying Context into ChatGPT and Claude

A familiar situation: the product team designs in one AI, runs an audit in another, and sometimes sends the same task through several services to compare the results. Each time, you have to explain the product, its constraints, and the decisions already made. If something changes in one chat, the others continue operating with an outdated version of reality.

The author of the Habr article proposes a simple solution: instead of copying context between platforms, keep it in one place. That is how AI Context Kit emerged, an open set of templates for a Git repository.

How it works

AI Context Kit is not a server, an MCP, or a standalone AI platform. It is a folder structure and a set of rules for working with it. At the root is PROJECTS.md, a project registry. Each project gets a folder containing three files:

• PROJECT.md: relatively stable information, including goals, audience, boundaries, and constraints.
• CONTEXT.md: the most dynamic section, recording decisions, results, open questions, and next steps.
• INSTRUCTIONS.md: rules for the AI itself, explaining how to use the documents, when to propose edits, and how to confirm them with the user.

The separation is deliberate. Goals may remain unchanged for months, while decisions are updated every week. This lets teams edit the current state without rewriting the description of the entire project.

A practical example

The demo creates a project called product-launch, a personal task-planning service. The interface concept is developed in ChatGPT by connecting the GitHub repository and specifying the relevant directory in Project Instructions. A decision is made: the main screen shows today's tasks, while the calendar remains an optional view. The decision is recorded in CONTEXT.md.

The audit is then conducted in Claude. The same repository is connected, and Claude is asked to read the latest CONTEXT.md. Claude sees the decision about the main screen and can work with it immediately, without a recap of the ChatGPT discussion. One important detail: the shared context contains neither the entire conversation nor every message, only the decisions and the information needed for subsequent stages.

Where the approach breaks down

The author is candid about the limitations. First, there is no automatic synchronization. Some services have built-in GitHub or MCP integrations, while others require files to be transferred manually. Second, Markdown instructions do not guarantee that a model will read every file or follow every rule, so teams should verify which context it actually used.

There is also a risk of conflicts: if two AIs edit CONTEXT.md at the same time, changes may be lost. The proposed safeguard is to check the file version before writing, for example by SHA, if the integration supports conditional writes. Edits are saved only after human approval.

Security deserves separate attention. Non-public projects require private repositories, and secrets must not be stored in them. Folders do not provide permission isolation, so projects with different access levels are better kept in separate repositories.

What to do with it

The idea is not revolutionary, but it is practical: context lives in Git, rather than in the memory of a particular service. You get change history, edit reviews, and vendor independence. For teams already using several models, it is an inexpensive way to eliminate manual copying.

You can start small: create a repository, move the project goals and list of accepted decisions into it, and connect it to your two main AI tools. The key condition for success is not technological but procedural discipline. If a decision is not added to CONTEXT.md, the next AI will not know about it. Context remains current only to the extent that you update it carefully.