You are training Agent Joe, a robot powered by language models, to navigate through mazes. Your job is to write prompts that tell Joe how to make decisions at each step. The better your prompt, the smarter Joe becomes!
You can write simple rules or complex templates. Use variables to access game state:
{vision.front}, {vision.right}, etc. - What Joe sees{previousActions} - Joe's recent moves and what he saw{direction}, {position} - Where Joe is and which way he's facingClick VARIABLES button for the full list and examples!
You can use these variables in your navigation prompt:
{vision} - Full vision (Front=X, Right=Y, Left=Z, Behind=W){vision.front} - What's in front (wall/empty/exit){vision.right} - What's to the right{vision.left} - What's to the left{vision.behind} - What's behind{direction} - Current facing direction (NORTH/SOUTH/EAST/WEST){position} - Current position as (x, y){position.x} - X coordinate{position.y} - Y coordinate{moveCount} - Total number of moves made{previousActions} - Last 5 moves with vision and actions taken<think>
I am at position {position}, facing {direction}.
I've made {moveCount} moves so far.
My current coordinates are: X={position.x}, Y={position.y}
Current vision: {vision}
Breaking it down:
- Front: {vision.front}
- Right: {vision.right}
- Left: {vision.left}
- Behind: {vision.behind}
Move history:
{previousActions}
</think>
Analyze the maze state and choose your next action.<think> I am at position (3, 5), facing NORTH. I've made 12 moves so far. My current coordinates are: X=3, Y=5 Current vision: Front=wall, Right=empty, Left=empty, Behind=empty Breaking it down: - Front: wall - Right: empty - Left: empty - Behind: empty Move history: Recent 3 moves (newest first): [1] Saw: Front=empty, Right=wall, Left=empty, Behind=wall → Action: FORWARD [2] Saw: Front=empty, Right=wall, Left=wall, Behind=empty → Action: FORWARD [3] Saw: Front=wall, Right=empty, Left=wall, Behind=empty → Action: RIGHT </think> Analyze the maze state and choose your next action.
Note: Variables are replaced with actual values before being sent to the model. You can write simple navigation rules or create complex templates using any combination of these variables.
This is where Agent Joe navigates!
Joe always starts at the top-left corner facing down (SOUTH).
The maze changes each time you press RESET.
Shows Joe's current state:
Shows what Joe sees around him:
The 3x3 grid represents Joe's view: