Prompt Files

Agent system prompts are stored as Markdown files in the prompts/ directory.

File Layout

prompts/
├── assistant.md       # General Assistant
├── mc-modder.md       # Minecraft Mod Developer
├── bug-fixer.md       # Bug Fixer
├── code-reviewer.md   # Code Reviewer
└── tester.md          # Test Assistant

There is no dedicated dispatcher prompt file — dispatcher-style agents (those with sub_agents) only write generic workflow rules and the sub-agent list is injected automatically (see below).

Writing Prompts

Prompt files are plain Markdown. The content is sent directly as the system message to the LLM.

You are a professional assistant. You can:
- Do something
- Do something else
Please answer concisely and clearly.

Tips

• Clearly list the agent's capabilities
• Specify desired output format
• Mention available tool types
• Keep it concise — avoid overly long prompts

Placeholders

Prompt files support ${placeholder} syntax for dynamic content injection. All placeholders are expanded at runtime when the system message is built.

Available Placeholders

Placeholder	Description	Example Output
${current_time}	Current time (HH:MM:SS)	14:30:00
${current_date}	Current date (YYYY-MM-DD)	2026-06-10
${current_datetime}	Current date and time	2026-06-10 14:30:00
${model_name}	Current LLM display name	DeepSeek V4 Flash
${model_id}	Current LLM model ID	deepseek-v4-flash
${agent_name}	Agent display name	General Assistant
${agent_id}	Agent config ID	assistant
${tools_overview}	All available tools (builtin + MCP)	See example below

Tool Overview Example

With ${tools_overview} in your prompt, the model sees a unified list of all tools — both builtin providers and external MCP servers from the mcps field:

- Unmodifiable File System (read_only): read_file, list_files, view_image, file_search, grep_search
- Modifiable File System (read_write): write_file, create_file, create_directory, replace_string_in_file, multi_replace_string_in_file
- Web Fetch (web): fetch_webpage
- TODO Tracking (todo): manage_todo_list
- Ask User (ask): ask_user
- Shell Command (shell): shell_exec
- ✓ GitHub (github-mcp, 26 tools: add_issue_comment, create_branch, create_issue, ...)
- ✓ Open WebSearch (open-websearch, 3 tools: search_web, fetch_page, ...)


### Recommended Usage

Always include `${tools_overview}` so the model knows which tools are available:

```markdown
You are a professional code reviewer. Current time: ${current_datetime}, running model: ${model_name}.

## Available Tools
${tools_overview}

## Responsibilities
- Review code quality, security, and performance
- Discover potential bugs and logic errors
- Provide improvement suggestions
- Check code style and best practices

Important

If you configure tools in mcps but forget to include ${tools_overview} in the prompt, the Agent card will show a ⚠️ warning badge.

Dispatcher Prompts

For dispatcher agents, only write the generic workflow rules in the prompt file. The sub-agent list is dynamically injected from agents.yml at runtime — no need to hard-code agent names:

You are an intelligent task dispatcher. Your job is to analyse user tasks,
decompose them into sub-tasks, and delegate to specialised sub-agents.

## Workflow
...

## Rules
1. Read the "Available Sub-Agents" list above and assign based on expertise
2. Each sub-task must be self-contained
3. Only assign sub-agents that are relevant
4. Aggregate results into a coherent final answer

Hot Reload

After editing prompt files, use the 🛠 Admin menu → Reload Local Config to hot-reload. New agent instances will use the updated prompts.

Related Config

Prompt file paths are specified in agents.yml via the agent_md_file field:

agents:
  - id: "my-agent"
    agent_md_file: "prompts/my-agent.md"