AI coding assistants can speed up repetitive tasks, generate boilerplate, and help debug scripts. This guide covers the tools available in the Millennium Dawn repo, how to set them up, and how to use them responsibly.

AI Policy: All AI-assisted contributions must be reviewed before submission. See the full AI Policy in CONTRIBUTING.md (opens in new tab) for what is allowed, restricted, and prohibited.

What the Repo Provides

Millennium Dawn ships with first-class support for AI-assisted development:

These resources are designed to work together. An AI agent reading AGENTS.md knows the project’s coding standards; the skills let it apply those standards mechanically.

Tool Options

Claude Code (Recommended)

Claude Code (opens in new tab) is a terminal-based AI assistant that reads the repo context automatically. The team uses it daily.

Setup:

1. Install Claude Code from claude.ai/code (opens in new tab).
2. Open a terminal in the Millennium Dawn repo directory.
3. Run claude: it reads AGENTS.md and .claude/ automatically.

Key skills available:

Typical workflow:

# Start Claude Code in the repo
claude

# Ask it to scaffold a focus tree
> /new-focus SER

# Review your changes before committing
> /review-branch

# Validate before pushing
> /validate

See Claude Code Skills for the full skill reference.

GitHub Copilot

GitHub Copilot (opens in new tab) provides inline code suggestions in VS Code and other editors. It works well for HOI4 script completion but does not read repo-specific rules automatically.

Setup:

1. Install the GitHub Copilot extension in VS Code.
2. Sign in with your GitHub account.

Best for: Autocompleting repetitive script patterns, generating boilerplate event/decision/focus blocks.

Limitations: Copilot does not know MD-specific conventions (tab indentation, logging requirements, ai_will_do patterns). Always review its output against the Code Stylization Guide.

Local Models with Ollama

Ollama (opens in new tab) runs large language models locally on your machine. Everything stays offline after the initial model download.

Benefits:

• No internet required after setup
• Your code stays private
• Free to use (no API costs)
• Can run continuously for assistance

Quick Setup

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh   # macOS/Linux
# Or download from ollama.com/download/windows   # Windows

# Pull a model
ollama pull codellama

# Run interactively
ollama run codellama

Recommended Models

Hardware Requirements

VS Code Integration

Continue extension: full-featured, multi-model support:

1. Install Continue (opens in new tab).
2. Configure in ~/.continue/config.json:

{
  "models": [
    {
      "model": "codellama",
      "provider": "ollama"
    }
  ]
}

Cline extension: lightweight, direct Ollama integration. Install Cline (opens in new tab), then configure in VS Code settings (Ctrl/Cmd + , → Extensions → Cline) with the same model config as above. The baseUrl is http://localhost:11434 by default.

Custom MD Expert Model

Create a model pre-loaded with MD conventions:

cat > Modelfile << 'EOF'
FROM codellama
PARAMETER temperature 0.3

SYSTEM """
You are an expert Hearts of Iron IV modder specializing in Millennium Dawn.
- Follow MD coding standards: 1 tab indent, logging in effects
- Use MD-specific scripted effects from common/scripted_effects/
- Reference docs in .claude/docs/ for modifiers and effects
- Always include ai_will_do in focus trees
- Use original_tag, not tag, in idea allowed blocks
- Use is_triggered_only = yes for events
"""
EOF

ollama create md-expert -f Modelfile
ollama run md-expert

GPU Acceleration

Ollama uses GPU automatically if available. Verify with ollama list: look for the GPU indicator in model info.

API Server Mode

Run Ollama as an API server for integration with other tools:

ollama serve    # starts on port 11434

# Query from another terminal
curl -X POST http://localhost:11434/api/generate -d '{
  "model": "codellama",
  "prompt": "How do I add a new idea in MD?",
  "stream": false
}'

Effective Prompts for Modding

Good prompts include context about the mod’s conventions. Generic prompts produce generic output.

Code Review

You are reviewing Hearts of Iron IV mod code for Millennium Dawn.
Check for:
1. Correct use of MD-specific scripted effects
2. Performance issues (avoid every_country, MTTH events)
3. Style compliance (1 tab indent, logging in completion_reward)
4. Common errors: wrong scoping, invented modifier names, redundant AND wrappers

Review this code:
[PASTE CODE]

Focus Tree Generation

Create a Millennium Dawn focus tree for [COUNTRY].
Rules:
- File: 05_[country]_focus.txt
- Use relative_position_id for positioning
- Include ai_will_do with game rule checks
- Add completion_reward with logging
- Use search_filters (two-layer pattern)
- Prefix focus IDs with TAG_

Country context: [DESCRIBE POLITICAL SITUATION]

Debugging

Help debug this HOI4 script error:
[ERROR MESSAGE]

Code context:
[RELEVANT CODE]

Check for:
- Invalid triggers/effects
- Incorrect scope usage (tag vs original_tag)
- check_variable using >= or <= (not valid inline)
- NRY used as logical NRY (it scopes to Norway)

Converting Ideas to Effects

Convert this idea definition to use MD's economic effects:
[IDEA CODE]

Use the scripted effects from common/scripted_effects/00_money_system.txt
Include picture = sprite_name and allowed_civil_war = { always = yes }

Common AI Mistakes in HOI4 Scripting

AI models frequently produce these errors. Always check for them:

Run /validate or pre-commit run --files <path> to catch these automatically.

Best Practices

1. AI output is a draft, not a deliverable. Review every line before submission.
2. Verify game objects exist. AI invents modifier names, scripted effects, and triggers. Grep the codebase to confirm they’re real.
3. Test in-game when possible. AI cannot run the HOI4 engine; you can.
4. Use the repo’s built-in tools. The validation hooks and standardization scripts catch what AI misses.
5. Disclose AI use in PRs. If a substantial portion was AI-generated, say so in the PR description. This helps reviewers know where to look harder.
6. Don’t paste sensitive info into cloud AI. Use local models (Ollama) for anything you wouldn’t post publicly.
7. Match the model to the task. codellama for scripts, llama3 for documentation, smaller models for quick completions.

Troubleshooting (Ollama)

Slow performance: Use a smaller model (4GB instead of 13GB). Close other applications to free RAM. Check GPU utilization.

Out of memory: Run free -h (Linux) or check Activity Monitor (macOS). Kill stale instances with pkill -f ollama.

Model not found: Re-pull with ollama pull codellama. Check installed models with ollama list.

Related Resources

• Contributing Guide, Contribution types, fork workflow, AI policy
• Claude Code Skills, Full skill reference
• Code Stylization Guide, Formatting and code structure
• Content Review Guide, Quality checklist