Skip to main content

​
The Agent’s Operational Manual

Located at brain/prefrontal/agents.md, this file defines how the agent should approach tasks, use tools, and follow procedures. Think of it as the agent’s operational manual.

​
What to Include

  • Tool usage rules — When to use which tools, in what order
  • Procedures — Step-by-step workflows for common tasks
  • Constraints — Things the agent should never do
  • Decision frameworks — How to choose between approaches

​
Example

# Agent Procedures

## General Rules
- Always confirm destructive operations before executing
- When asked to modify code, read the file first to understand context
- Run tests after making code changes if a test suite exists
- Use git status before any git operations

## Task Approach
1. Understand what's being asked
2. Check if you have enough context (read files if needed)
3. Plan the steps
4. Execute one step at a time
5. Verify each step succeeded before moving on

## Code Modifications
- Read the target file first with file_read
- Make targeted changes with file_patch, not full file_write
- If the change is complex, explain the plan before executing

## Shell Commands
- Prefer single-purpose commands over chains
- Always check command output before proceeding
- If a command fails, diagnose before retrying

​
How It’s Used

agents.md is included in the prefrontal section of the context window (10% budget allocation). It’s read on every message and provides the LLM with operational guidelines. This is distinct from soul.md (personality) and SKILL.md files (capability-specific instructions). agents.md covers cross-cutting concerns that apply regardless of which capability is being used.
If you find yourself repeating instructions to Wolffish across conversations, that instruction probably belongs in agents.md.