← Back to comparison

Documentation

The easiest way to tell the LLM how to use libraries, APIs, and other resources.

Filesystem

You can create an AGENTS.md file in the root of your project to tell the LLM how to use the agents.

Your agent will include this file in the context for each session. It acts as the system context for the project. You can declare anything in it: how to call local tools, what steps to create in plan mode, which files to generate, or any instructions you want.

More info: https://agents.md/

Note: Claude Code doesn’t support this feature. Instead, it uses a different filename, CLAUDE.md.

Note: Cursor also supports a .cursor/rules/ folder, which is a similar feature. More info: https://cursor.com/docs/context/rules

Tree

If a file isn’t part of the context, you don’t need its documentation included either.

If your agent adds a file to the context, it will also include any documentation files from the same directory.

For example:

root
├── lib1
│   ├── Claude.md
│   └── lib1.source
├── lib2
│   ├── Claude.md
│   └── lib2.source
└── code.source

If root/code.source only uses root/lib1/lib1.source, then only root/lib1/Claude.md is added to the context. root/lib2/Claude.md is skipped.

Only this agents supports this feature:

Multi-file

Put docs for the same folder in separate files.

Sometimes it is better to split the documentation into multiple files. For example:

root
├── .cursor
│   └── rules
│       └── deploy.mdc // describes how to deploy the project
│       └── lint.mdc // describes how to lint the project
│       └── test.mdc // describes how to test the project
├── lib1
│   ├── .cursor
│   │   └── rules
│   │       └── wasm.mdc // describes how to use the wasm library
│   │       └── js.mdc // describes how to use the js library
│   └── lib1.source
├── lib2
│   ├── Claude.md
│   └── lib2.source
└── code.source

If root/code.source uses root/lib1/lib1.source, the system context will be like:

<root/.cursor/rules/deploy.mdc>
<root/.cursor/rules/lint.mdc>
<root/.cursor/rules/test.mdc>

<root/lib1/.cursor/rules/wasm.mdc>
<root/lib1/.cursor/rules/js.mdc>

root/lib2/Claude.md is skipped because it is not part of imported files.

Only this agents supports this feature:

LLMS-txt

Let the author of the service provide the docs

Our code relies on many external sources, such as libraries and APIs. Only the service authors truly know how to use them most effectively.

The community has introduced a useful format: llms-txt. For example:

However, these files are typically very large and cannot be fully loaded into an LLM’s context. Some agents natively support selective inclusion of these files in context.

One approach is RAG (https://en.wikipedia.org/wiki/Retrieval-augmented_generation), which allows the agent to include only the relevant sections from the entire llms-full.txt file as needed.

Only this agents supports this feature natively:

Auto-merge

Let the agent pick the right docs for you.

There are many llms-txt files available across the web—for example, https://github.com/SecretiveShell/Awesome-llms-txt. Users often aren’t aware that some of these files exist. Ideally, the agent should be proactive and suggest helpful files. For example:

# User request:
Add a new feature to the 'example.astro' file.

The agent should suggest including the Astro llms-txt file in the context.

# Agent response:
I see you're using the Astro framework. Would you like to add their llms-txt file to your context?

None of the tested agents supports this feature.

Partial

Let the agent decide what part of you local docs should be included in the context.

Sometimes you have a lot of local docs, but you don’t need all of them. For example:

root
├── .cursor
│   └── rules
│       └── deploy.mdc // describes how to deploy the project
│       └── lint.mdc // describes how to lint the project
│       └── test.mdc // describes how to test the project

If you want only to tune the deployment process, you don’t need to include the lint and test docs into the context.

Only this agents supports this feature: