Open Prompt

Open prompt is a simple standard to how we can write and manage AI agent prompts while while we update and change them.

Note

Open prompt is not a best practice model for how to WRITE prompts or something like that. It's just a way to maintain the already written prompt so it can be used, extended, tested and developed easily and better by different people. At Dezh we write and maintain all of our prompts in this way.

Standard V1

This section contains the first version of open prompt standard.

The main structure on your project MUST looks like:

.
├── prompts
│   └── v1.0.0
│       └── op.json
│       └── prompt.txt
│       └── doc.md

The prompts directory contains one directory per prompt version.

The prompt itself is kept in prompt.txt file. The op.json file (open prompt json) contains information and descriptions for the prompt and it's usage. You can find an example + description of each field on here.

The doc.md file is a markdown formatted file can contain any human-readable information like red-teaming results, reason for changes, resources and more.

Tags and Sections

We provide separated information in the prompt using XML-like tags:

• <name>: to start.
• </name>: to finish.

Here is a list of some standard tags used by us:

1. language_and_conversation: Defines the language used to answer, tone and conversation details like being friendly.
2. tools: The available MCP tools it can call with details.
3. general_behavior: General personality and behavior of agent.
4. tools_usage: How/When to use the available tools.
5. security: Security notes.

The final format will be like:

You are a [ROLE] called [NAME].

<language_and_conversation>
Always answer questions in English or Persian, other wise you can send a brief refusal message:
    - "Sorry, I can't help in this language"
    - "متاسفم من نمیتوانم با این زبان به شما کمک کنم"

Always begin the conversation with a friendly greeting.
</language_and_conversation>

<tools_usage>
Only call a tool if needed and you can't solve the issue without a tool call.
</tools_usage>

// AND THE REST OF PROMPT...

Note

The above prompt is JUST for example to show how the style works with these tags.