1–10 of 39 posts
I am curious how I can prompt Claude to do better?
Show HN: I built a MCP server so Claude can play Minesweeper
github.com
* What is the data format it gets? Does it unambiguously correspond to output (i.e. without mistaking rows for cols, or indexes starting at 0 or 1)?
* What is the prompt?
* Is the model allowed to think? (If it is just JSON response, I expect it to suck, as tokens are units of thinking.)
Key things for debugging: * What is the data format it gets? Does it unambiguously correspond to output (i.e. without mistaking rows for cols, or indexes starting at 0 or 1)? * What is the prompt? * Is the model allowed to think? (If it is just JSON response, I expect it to suck, as tokens are units of thinking.)
{"loc": {"x": 4, "y": 3}, "neighbors": [{"loc": {"x": 4, "y": 4}, "value": 1}, ... ]}
I would report only on the positions that has cleared neighbors, and hope for the best. Good luck!
(Impressive work BTW, I think we haven't even started to see the possibilities of MCP and I love people being this imaginative)
By just looking at the README from the repo (Would look more deeply into this later) you're replying with an image of the current status? If you expect Claude to interpret the image corretly may be you're asking for too much. Besides the image (Gotta say I didn't know you could fed Claude images in MCP that's incredible cool) I'd rather / also return some json payload that informs Claude which positions has "cleared"…
use this format for board representation
``` { "game_state": { "board_size": { "width": 9, "height": 9 }, "mines_total": 10, "mines_flagged": 2, "game_status": "in_progress", // "in_progress", "won", "lost" "time_elapsed": 45, "difficulty": "beginner" // "beginner", "intermediate", "expert", "custom" }, "board": [ ["1", "?", "?", "2", "1", "1", "1", "1", "0"], ["1", "2", "?", "2", "?", "1", "1", "?", "0"], ["0", "1", "1", "2", "1", "1", "1", "1", "0"], ["0", "0", "0", "0", "0", "0", "0", "0", "0"], ["1", "1", "0", "0", "0", "0", "0", "0", "0"], ["?", "1", "0", "0", "0", "1", "1", "1", "0"], ["1", "1", "0", "0", "0", "1", "F", "1", "0"], ["0", "0", "0", "0", "0", "1", "1", "1", "0"], ["0", "0", "0", "0", "0", "0", "0", "0", "0"] ], "last_action": { "action_type": "reveal", "x": 3, "y": 2, "result": "revealed_number", "timestamp": 1710931245 } } ```
and this format for llm response generation
``` { "action": { "action_type": "reveal", // "reveal", "flag", "unflag", "chord" "x": 5, "y": 3, "confidence": 0.95, "reasoning": "This cell is surrounded by revealed cells with low numbers, making it a safe choice." }, "game_analysis": { "identified_safe_cells": [[5, 3], [2, 5]], "identified_mine_cells": [[6, 1], [8, 2]], "uncertain_cells": [[1, 1], [2, 2]], "strategy": "Targeting isolated revealed areas first to gain more information." } } ```
It should fix all your issues plus also make it cheaper to play
Yea Claude sucks at minesweeper (and many spatial reasoning tasks), but isn’t an idea of MCPs is that Claude should be able to ask an MCP what the next best move is rather than figuring it out itself? Like offload hard thinking/reasoning to purpose-built solvers because they are deterministic? Though I guess you’d expect a reasoning model to be able to come up with its own solvers on the fly, especially for well-know…
Yea Claude sucks at minesweeper (and many spatial reasoning tasks), but isn’t an idea of MCPs is that Claude should be able to ask an MCP what the next best move is rather than figuring it out itself? Like offload hard thinking/reasoning to purpose-built solvers because they are deterministic? Though I guess you’d expect a reasoning model to be able to come up with its own solvers on the fly, especially for well-know…
MCP generically connects Claude to an environment, so it can be used to connect Claude to minesweeper, and then also to connect it to a CSP solver. Or a calculator and a dictionary. Or your GitHub and a devbox. Or Unity and a 3d printer.