Scrabble Challenge

Addressing the challenge posed in this tweet by Karl Muth.

Claude Opus one-shot a Python script that searches a Scrabble dictionary to build a complete, valid board and renders it as an image.

Why this is hard

Producing a valid Scrabble board requires simultaneously honoring the dictionary, tile inventory, point values, the 15×15 grid with its specific premium-square layout, and the constraint that every word connects through shared letters. Frontier image-generation models do not get this right. The example_failed/ directory contains attempts from ChatGPT and Gemini that illustrate the failure modes:

• Gemini: non-words next to real ones; rectangular cells instead of squares; premium squares in the wrong positions; too many cells and too many tiles; inconsistent and incorrect point values; disconnected words; and the model rendered the prompt's color shorthand ("B" for beige, "LB" for light blue) as literal tiles — the board has 24 "B" tiles and two "LB" tiles.
• ChatGPT: a tile placed in nearly every non-premium square, producing far too many tiles and failing to spell valid English words.

Solving the problem properly is a search/constraint problem, not a rendering one — which is why the approach here is a Python script that searches the dictionary and then draws the result, rather than asking a model to draw the board directly.

Contents

• generate_scrabble_board.py — the script
• prompt.txt — the prompt used
• metrics.txt — run metrics
• dictionaries/ — TWL06 word list
• output/ — rendered board image
• example_failed/ — failed attempts from popular image-generation models