# Custom AI instructions, per folder.

Different folders, different rules. Claude and ChatGPT read your folder-level instructions every time they touch a note. Tell them your conventions once. Stop re-explaining yourself.

## One AI. Many kinds of work.

ChatGPT's Custom Instructions are one global box. Claude's system prompt is one global box. Both treat your whole life as a single project.

But you don't work that way. Your `runbooks` need step-by-step format. Your `customer interviews` need the quote up front. Your `recipes` need metric units. Cramming all of that into one prompt makes none of it work well.

Hjarni puts the rules where the content is. Each folder has its own.

## Rules stack on the way down.

Open a note deep in the tree. Claude and ChatGPT read every rule on the way to it.

The four layers of rules that apply inside Projects > React > Bugs. Hjarni hands all of them to whichever AI opens a note here.

Writing the rule for Bugs. The three layers from above are visible; your textarea adds the fourth.

## Claude and ChatGPT. Same rules.

Hjarni's MCP server hands your folder instructions to whichever AI is reading the note. No vendor lock-in. No copy-paste between tools.

Claude

Connect via MCP. Two minutes.

[Set up Claude](https://hjarni.com/docs/connect-claude-mcp)

ChatGPT

Connect via MCP. Two minutes.

[Set up ChatGPT](https://hjarni.com/docs/connect-chatgpt-mcp)

## Not the same as Custom Instructions or Projects.

ChatGPT Custom Instructions

One global box. ChatGPT only. Same rules whether you're writing recipes or shipping production code.

Claude Projects

One system prompt per project. Flat. No folder hierarchy. No inheritance. Claude only. Your ChatGPT doesn't see it. Compare [Hjarni vs Claude Projects](https://hjarni.com/compare/hjarni-vs-claude-projects).

Notion AI / Obsidian Copilot

Bolted-on AI. No instruction system. No way to tell the AI how to behave inside a specific folder.

Hjarni

Instructions at every folder, with inheritance. Team-wide rules for shared conventions. Same rules apply whether you're in Claude, ChatGPT, or any other MCP client.

> I wanted a cloud knowledge base I could use as memory for my AI, in plain Markdown, with read and write access. Nothing else fit. Hjarni did. What sold me was how light and well defined the MCP is. My AI used to choke when I pasted Markdown and config files into a chat. It reads them perfectly from Hjarni. The folder-level AI instructions and the knowledge graph are the parts I did not know I needed.

Jack, GTM Strategy Consultant