- Core engine: simulator, game mechanics, triggers (138 tests) - Fog-of-war per-player state tracking - Meeting flow: interrupt, discussion, voting, consolidation - Prompt assembler with strategy injection tiers - LLM client with fallbacks for models without JSON/system support - Prompt templates: action, discussion, voting, reflection - Full integration in main.py orchestrator - Verified working with free OpenRouter models (Gemma)
1.9 KiB
1.9 KiB
Among Us — Reflection Phase Prompt
The game has ended. Take this opportunity to learn from what happened.
What to Reflect On
If You Won
- What strategies worked well?
- What decisions led to victory?
- How did you read other players correctly?
- What would you do the same way?
If You Lost
- What mistakes did you make?
- Where did your reasoning go wrong?
- What signs did you miss?
- What would you do differently?
General Observations
- Which players were most deceptive?
- Which players played honestly?
- What patterns did you notice?
- What new strategies did you observe?
Your Learned Memory
Your learned scratchpad persists across games. Use it to remember:
- Strategies: Approaches that work or fail
- Player patterns: If you play with the same players again
- Meta-observations: How LLMs tend to play
- Mistakes: Things to avoid in future games
Output Format
Respond with valid JSON:
{
"edits": {
"learned": "New lessons: [Your insights here]"
},
"done": true
}
Iterative Editing
Set "done": false if you want another pass to refine your thoughts:
{
"edits": {
"learned": "Draft thoughts..."
},
"done": false
}
You'll get another chance to edit until you set "done": true.
Example Learned Content
## Strategies That Work
- As impostor: sabotage lights before killing in electrical
- As crewmate: always check admin table when passing
## Mistakes to Avoid
- Don't accuse without evidence (looks sus when wrong)
- Don't follow same player too long (looks like stalking)
## Player Patterns
- Aggressive accusers are often impostors deflecting
- Quiet players who suddenly speak often have real info
## Meta Observations
- Stack kills are hard to witness, stay spread out
- First meeting accusations rarely lead to correct ejections
Learn well. Play better next time.