wtf is pictionary?

Pictionary is a playful tool built around a quirk of newer Claude models. Those models no longer accept the temperature setting, a parameter that used to let people control how random or varied an AI's answers were. Instead of asking for that control back through a prompt, this project takes a different route: it turns your text prompt into an image, then blurs that image by a chosen amount before feeding it to the model. Because the model has to read blurry text from a picture instead of clean text, its occasional misreadings act like a stand in for the randomness the temperature setting used to provide. The author tested this and found that turning up the blur level made Claude's answers to the same question noticeably more varied, roughly three times more so at higher blur than at none. To use it, you install the tool globally, add its accompanying skill to Claude Code, and then run a command that packs a text file into a blurred image at whatever blur level you want. Inside Claude Code, you then have the model read that image file instead of the original text file. Beyond the temperature trick, the tool has a second, more practical side effect: images are billed differently than text by the model, and packing enough text into a single image page can end up cheaper than sending the equivalent plain text, especially at higher pixel density settings, though a dry run mode lets you check whether this actually saves money for a specific file before committing to it. The author is upfront that this is a small, half-joking project rather than a serious, production-ready sampling system, and warns against using it for text that must be read back exactly, like ID numbers, hashes, or code meant to be edited, since misreading blurry text is part of how it works. It has one runtime dependency and is released under the MIT license.