Snake Lab Prompts

These prompts are used by the current Snake Lab single-and-pair optimization loop, using text-only messages, including feedback sent when a submission needs correcting. The golden configuration is the current baseline used to evaluate new proposals.

Prompt name Description
ProjectContext Opens every conversation with the tutorial context and the LLM’s responsibility for choosing parameters.
FirstContact Introduces Patrick Loeber’s AI Snake Game, sets the high-score objective, and encourages exploring apparently poor choices to gather data and discover higher plateaus.
FirstContactSingle Requests an untested value for the assigned parameter using submit_single_value({"value": number}).
Comparison Shows only the assigned parameter’s golden value, high score, and comparable history on the current and earlier seeds. Used from the first conversation onward.
FirstContactEpsilonPair Requests an initial/decay proposal through submit_pair_values, on each epsilon turn.
ComparisonEpsilonPair Shows the golden epsilon pair and sorted comparable scores across seeds.
FirstContactRewardPair Requests a closer/further distance-reward proposal through submit_pair_values, on each reward turn.
ComparisonRewardPair Shows golden rewards and the full 7-by-7 grid of comparable scores across seeds.
ComparisonSingle Requests the next untested value for the assigned parameter.
InvalidValue Explains why a submission was invalid and asks the model to choose a legal value from the JSON specification and resubmit.
NoDupes Explains that a proposed configuration matches the golden configuration or an existing simulation, and asks the model to submit a different value.

This list follows SnakeLabLoop.py, ToolConversation.py, and SubmitSingleValueHandler.py in ax3l/app/snakelab/. The round robin visits five singles, then epsilon and reward pairs.

Requests require one tool call through the API’s tool metadata. A response without the expected tool call ends the conversation with a logged error. Rejected values still receive validation feedback within the same conversation.

LossPlot and ComparisonPlot are no longer used by the optimization flow.

The MCP tool description supplies only the assigned parameter’s meaning and schema rules. The full golden configuration and other parameters’ histories are not sent.