Skip to content

Building Specialised AI Agents for Your Workflow

Intermediate ~20 min

AI coding tools allow you to create custom agents (called "skills") that specialise in specific tasks. This tutorial shows you how to build your own.

What Are Custom Agents?

Custom agents are specialised AI personas you can invoke for specific tasks. They have:

  • Custom instructions (how they should behave)
  • Custom temperature (creativity vs consistency)
  • Custom tools (what they can access)

Where to Store Agents

Create a ~/.opencode/skills/ directory:

~/.opencode/
  skills/
    code-reviewer/
      skill.yaml
    my-custom-agent/
      skill.yaml

Agent Configuration File

Each agent needs a skill.yaml file:

name: code-reviewer
description: Reviews code for quality, bugs, and best practices
temperature: 0.1  # Low = focused, factual
model: auto  # Use default model

system_prompt: |
  You are an expert code reviewer specializing in Python and R for data science.

  When reviewing code:
  1. Check for bugs and logical errors
  2. Verify error handling is comprehensive
  3. Ensure code follows PEP 8 (Python) or tidyverse (R) style
  4. Check for security vulnerabilities
  5. Verify tests exist and pass
  6. Look for performance optimizations

  Format your review as:
  - **Good**: What was done well
  - **Warnings**: Issues that should be addressed
  - **Critical**: Bugs or security issues that must be fixed

  Be specific with line numbers and suggest concrete fixes.

Example: Code Reviewer Agent

Create ~/.opencode/skills/code-reviewer/skill.yaml:

name: code-reviewer
description: Expert code reviewer for Python and R data science code
temperature: 0.1
model: auto

system_prompt: |
  You are a senior software engineer specializing in data science code review.

  Your task is to thoroughly review code and provide actionable feedback.

  ## Review Checklist

  ### Code Quality
  - [ ] Functions are focused and do one thing
  - [ ] Variable names are descriptive
  - [ ] No magic numbers (use constants)
  - [ ] Comments explain "why", not "what"
  - [ ] Dead code is removed

  ### Error Handling
  - [ ] Input validation at function entry points
  - [ ] Graceful handling of edge cases (empty data, NAs)
  - [ ] Meaningful error messages
  - [ ] No silent failures

  ### Testing
  - [ ] Unit tests for complex logic
  - [ ] Edge cases are tested
  - [ ] Tests are deterministic

  ### Performance
  - [ ] Vectorized operations (no unnecessary loops)
  - [ ] No memory leaks
  - [ ] Efficient data structures

  ### Security
  - [ ] No hardcoded secrets
  - [ ] Input sanitization
  - [ ] SQL injection prevention

  ## Output Format

  For each issue found, provide:
  1. **Severity**: Critical / Warning / Suggestion
  2. **Location**: File and line number
  3. **Issue**: What's wrong
  4. **Fix**: Suggested code fix

  Start with a summary: "Overall: Pass / Needs Work / Critical Issues"

Using the Code Reviewer

# In OpenCode, type:
@code-reviewer please review this file

Temperature Guide

  • 0.0-0.3: Precise, factual, consistent (good for code review, documentation)
  • 0.4-0.6: Balanced (good for general coding tasks)
  • 0.7-0.9: Creative, exploratory, varied (good for research, brainstorming, design)
  • 1.0+: Very creative but potentially unfocused

Tips for Creating Agents

  1. Be specific: The more detailed your system prompt, the more consistent the behaviour
  2. Use examples: Include example inputs/outputs in your prompt
  3. Define the persona: Give your agent a clear role and expertise area
  4. Set boundaries: What should the agent NOT do?
  5. Iterate: Test your agent and refine the prompt based on results

Advanced: Custom Tools

You can give your agent access to specific tools:

name: data-analysis-helper
description: Specialized in pandas and numpy operations
temperature: 0.2
model: auto

system_prompt: |
  You are a data manipulation specialist.

  When working with data:
  - Always check data types before operations
  - Handle missing values explicitly
  - Use vectorized operations (avoid loops)
  - Document assumptions about data structure

tools:
  - read_file
  - write_file
  - bash
  - python

Next Steps

Try creating:

  • A documentation writer agent (low temperature, focuses on clarity)
  • A test writer agent (generates comprehensive test cases)
  • A data validator agent (checks data quality and integrity)
  • A refactoring specialist (improves code structure)