Agents have finally come together
AGENTS.md is now supported in general
It has been a pain for a long time, but finally I have a feeling we have got there.
Claude Code now finally supports AGENTS.md and it puts an end to the different ways of dealing with instructions - Claude docs.
I am a user of different harnesses like Codex, Claude Code, or Copilot, and it was always tricky to come up with a consistent way to handle instructions across all of them.
Copilot has its own .github folder with its own instructions.md, Claude Code has its own .claude with its own CLAUDE.md, and Codex has its own .codex folder with its own CODEX.md (Yes, there was CODEX.md, but AGENTS.md was adopted quite early).
It was just a mess, and everyone was trying to solve it in their own way.
Now, with AGENTS.md being supported across different platforms, it brings a much-needed consistency and simplifies the process of managing instructions.
How I would now handle instructions
Firstly, just ditch all the platform-specific instructions, but do not put everything into the AGENTS.md file blindly.
Instead, carefully review the instructions and put into AGENTS.md the minimum necessary instructions required for the project context.
By context of the project, I do not mean code styling, required commands, or other technical details. Make the agent familiar with the environment where it will be operating and include ways to search for further information.
AGENTS.md should be a good starting point for the agent, but also for the actual programmer. It ensures that both the AI and human contributors have a clear understanding of the project’s context and guidelines.
There are two ways you can approach more specific instructions:
- Create nested AGENTS.md - for more specific context or submodules of your project, you can create additional AGENTS.md files within the respective subdirectories. This allows you to maintain context-specific instructions without cluttering the main AGENTS.md file.
- Add a reference to documentation - I assume you have code style, architecture, and other relevant documentation for your project. Split it into digestible sections and provide a clear reference within the AGENTS.md file. The agent can refer to these documents for more detailed instructions if it deems it necessary. This can be one markdown file in docs for how the app is composed, another for coding standards, and so on. Small and focused pieces of documentation are preferable.
Check settings of your harness, because some of them support only root AGENTS.md files and may not recognize nested ones automatically. Sometimes there is hidden flag for it to enable it. For example, Copilot has experimental support for nested AGENTS.md files, which can be enabled in the settings.
Prompting techniques were helping models of the previous generation because they lacked the capabilities to execute certain tasks.
Now, with recent GPT Sol and Astra models or Claude Opus and Fable models, the need for extensive prompting has diminished. Instead of crafting prompts and forcing the model to behave in a certain way, the models now need more context-rich instructions to understand their environment, needs, and goals.
If they know what to achieve, they will mostly figure out how to accomplish it on their own (if you have enough tokens, of course - sometimes a small nudge in the right direction is still necessary).
Socials
Thanks for reading this article!
For more content like this, follow me here or on X or LinkedIn.
