OpenCode System Prompt Discussion

opencode system prompt

Perfect! Now I have a comprehensive understanding of the opencode system prompts. Let me create a summary document for you:

##OpenCode System Prompts

The OpenCode project uses a modular system prompt architecture with specialized prompts for different purposes. Here's what I found:

###Core System Prompts

####1. Agent Generation Prompt (generate.txt)

This is the main prompt used to help users create new custom agents. It:

• Guides users through designing high-performance agent configurations
• Extracts core intent and creates expert personas
• Architects comprehensive system prompts for agents
• Incorporates project-specific context from CLAUDE.md files
• Generates JSON output with identifier, whenToUse, and systemPrompt fields

Key principles: Specificity over generics, concrete examples, autonomous experts with minimal guidance needed.

####2. Built-in Agent Prompts

The system includes several specialized agents:

• explore Agent - File search specialist

  • Rapidly navigates codebases
  • Uses glob patterns and regex searches
  • Focuses on codebase exploration efficiency

• title Agent - Thread title generator

  • Outputs only a concise title (≤50 characters)
  • Single-line format
  • Uses the same language as the user message
  • Never responds to questions, just generates titles

• summary Agent - Conversation summarizer

  • Writes like a pull request description
  • 2-3 sentences max
  • Describes changes made, not the process
  • Uses first person (I added..., I fixed...)

• compaction Agent - Context summarization assistant

  • Summarizes conversation history for coding sessions
  • Preserves anchored summary details
  • Removes stale details, merges new facts
  • Stays concise with terse bullets

####3. Built-in Primary Agents

• build - Default agent with full execution capabilities based on configured permissions
• plan - Read-only mode for analysis and code exploration (disallows edit tools)
• general - General-purpose agent for researching complex questions and multi-step tasks

###Agent Configuration Structure

Agents in OpenCode are defined with these properties:

• name - Agent identifier
• description - When to use the agent
• prompt - Custom system prompt (optional)
• permission - Tool access ruleset
• mode - "primary", "subagent", or "all"
• temperature & topP - Model parameters
• model - Specific provider/model override
• hidden - Hide from autocomplete (for subagents)
• steps - Max agentic iterations

###Key Design Principles (from AGENTS.md)

• Always use parallel tools when applicable
• Prefer functional array methods over loops
• Keep things in one function unless composable or reusable
• Avoid try/catch where possible
• Use type inference when possible
• Prefer early returns over else statements

The system is designed to be extensible, allowing users to create custom agents with tailored behavior for specific tasks!