Show HN: I over-engineered a dingbats style puzzle game with AI agents(thingbat.today)
thingbat.today
Show HN: I over-engineered a dingbats style puzzle game with AI agents
https://thingbat.today/
1 comments
I've tried generating some dingbat puzzles with Gemini, they are comically bad most of the time with no subtlety. Even for existing dingbat phrases it didn't make good ones, which I would have expected it to have been able to just copy prior examples.
Same with cryptic crosswords, not very good. Sure there's some strategy that would work like providing some tactics to use.
Same with cryptic crosswords, not very good. Sure there's some strategy that would work like providing some tactics to use.
i wanted to play it enough to build it - but i quickly realised it's not very fun to play your own puzzles.
so i built a complex, ai agent flow to autonomously spoon feed me puzzles, they were a mix of:
1. incredibly obvious (e.g. answer written straight in plain text) 2. completely incomprehensible (e.g. no idea how you could get the answer from the image)
so i "improved" the system by adding more agents to check the work of the agents and rate the subtlety and solvability of puzzles using a range of complex heuristics.
which worked well since it filtered out 100% of the ai generated puzzles, then i added some feedback loops from these agents to the puzzle generation agent...
which made it very expensive to run.
then, a friend of mine told me how much he also loves these types of puzzles, so we deleted (most) of the ai agents - and we use more basic tooling to set each-other puzzles each day.
have fun.
footnote: [1] the dailydingbats IG is pretty good though, check it out if you like dingbats