Working with the Agent

The Parallax agent is not a search engine — it's a collaborator. The more you treat it like a skilled developer who knows your game, the better results you'll get.

Prompting patterns that work well

Describe the behaviour, not the code

The agent infers the implementation. You describe the experience.

Scope your requests

One well-defined change per prompt is faster and easier to review than a large batch:

• ✅ "Add wall-jumping to the player"
• ✅ "Make enemies patrol left-right between two points"
• ⚠️ "Add wall-jumping, enemy patrol, and a level transition screen" — split this into three prompts

Reference your own files

Look at how I handle player movement in player.lua and apply the same pattern to the enemy

The agent understands your codebase — referencing existing files produces more consistent output.

Iterate with feedback

That's almost right — but the jump arc feels floaty. 
Can you reduce the jump height by 20% and add a small squash animation on landing?

The agent keeps context between messages in the same project session.

What the agent knows

The agent is grounded in:

1. Your project files — always in context
2. Love2D 11.5 API — all callbacks, modules, and functions
3. This documentation — best practices, mental models, common patterns
4. Your .parallax/context.json — your game's rules and conventions

What the agent doesn't do

• It won't make large architectural changes without explaining the trade-offs
• It won't silently delete files — all file changes are shown as diffs before writing
• It won't invent Love2D APIs that don't exist (if it's uncertain, it'll say so)

Reviewing diffs

Every agent response that modifies code includes a diff. Take 30 seconds to read it — it's the fastest way to build intuition for Love2D patterns and catch anything unexpected.