Hitzai by ejtorre

The player is a king's constable ordered to investigate the disappearance of the blacksmith Moran in the village of Valdegris. They arrive at the village at dusk with the mission to find Moran and uncover the truth about the Shadow Griffin.

Hitzai is a turn-based narrative game played entirely through a chat interface. The player must achieve the story goal chating with the AI narrator who has the game rules, world, goals, etc. The player has a limited number of turns to complete the mission.

HOW IT WORKS

Read the opening of the story.

Type your action, question, response, etc.

The AI narrator processes your action and returns a response.

You have a limited numbers of turns to achieve mission's goal.

FEATURES

AI narrator that responds to free-form text input.

Multilingual: Currently there are six available languages:Spanish, English, French, German, Italian and Portuguese.

Free to play. Runs in your browser: https://www.hitzai.com

Play online

Comment ejtorre

12 Apr 2026

As I have many visits to my game coming from this web, I think that I have to publish here recent important changes to the game:

I was using the Groq API with the model meta-llama/llama-4-scout-17b-16e-instruct but that model was not allowed in EU, so I changed by the model to qwen/qwen3-32b. The model was good but Groq API only allows 6K tokens per minute which was not enough for even one single game because the input grows as the game advance. Then I had to find an alternative. After several comparisons, I decided to choose Google API and the model gemini-3.1-flash-lite-preview. Currently, the free version of the Google API for the model gemini-3.1-flash-lite-preview allows 15 RPM (request per minute) and 250K TPM (tokens per minute) which is enough for the current use of the game. Unfortunately, the API calls are a bit slower than meta-llama/llama-4-scout-17b-16e-instruct with Groq API.

Comment ejtorre

05 Apr 2026

On April 2nd, I updated the game and one of the main changes was the game over screen, where I added the LLM's final response which is the actual conclusion of the story. I’d love to share a screenshot here, but it doesn't seem to be possible in these comments.

Regarding how the LLM 'kills' the player, it really depends on what the player writes. In the system prompts, I haven't hard-coded specific 'if X happens, then kill player. I’ve provided clear instructions on the win and loss conditions, and I've guided the LLM to include the words 'victory' or 'defeat' in its response when those conditions are met. Once that happens, the code takes over to trigger the end-game screen.

Comment 2theleft

01 Apr 2026

Thanks @ejtorre I've seen the end modal for both win and lose and it's good. It would be handy if the Log button was available on it.
You know, I hadn't actually noticed there was a final response I hadn't been seeing, I just accepted that I'd abruptly run out of road. It is interesting that it can generate the ending, and that it can shepherd the player towards a hard turn limit - or is it my imagination that it can do that? I get the strongest feeling it doesn't like me dawdling or digressing too much in the early turns, and conversely sends in interruptors if I try and bring on the denoument too soon in the later turns.
If playing Copilot or Gemini directly on line, I don't sense anything like that.

If it's a trade secret, fair enough, but how do you get it to kill the player at all? I'm usually quite good at working on the sycophancy in these things, but your one can turn round and bite the user!

Comment ejtorre

30 Mar 2026

Thank you for your positive review. I did indeed decide to limit the number of turns and set clear objectives to avoid the game becoming simply an endless story. My idea was to create a sense of game rather than just a chat with an LLM (ChatGPT, Claude, Gemini, Groq, etc). On the other hand, it is worth noting that running LLM inferences currently consumes significant resources so limiting the number of turns also limits token consumption.

I am still working on frontend improvements. For example, yesterday I replaced the game's end modal with a new one that shows the last message sent by the AI narrator. Previously it was hidden once the game result screen appeared, meaning the player never got to read the end of the story.

Review 2theleft

29 Mar 2026

This is the first game I've tried with an AI-based interface, so I'm giving it 5 because of wow!

I really like the freedom it gives you, the player. There's no more struggling to make yourself understood. The world fleshes itself out as you drill down to arbitrary depths of detail.

The Shadow Griffin (the thing I think you are meant to sort out) is quite an elusive bird, in the sense that it will be quite different from one run of the game to the next. Very little of what you learn can be carried forward to next time. This could make it an endless source of entertainment for some. Or it could undermine the reality of the challenge.

There is a strict cap on the number of turns you can take. I suspect this is to prevent the drift of the brief becoming apparent. LLMs doing this sort of thing tend to drift until they get caprtured in a tight pattern.

It will be very interesting to watch how this medium develops.