[Feature Request] Un-hardcode AIPersonaity and AICommandsAndQueries

[dehen]

This is probably already planned, but having a text file like the correction-dictionary where you could edit/add personalities, or add a phrasing for a particular command would be very nice.

Having access to these files in plain-text would also give the users another way to discover what sorts of commands are available.

Thanks for this program. I just started using it today and it's lots of fun.

[stone-alex]

Hey, @dehen good questions! And yeah, it's something we've thought about.

Personalities

The short version is: we deliberately keep the "personality" part of the AI's instructions locked down and not user-editable.
Why? Because the AI works with a fixed "budget" of tokens (think of them as words/letters it can read at once). Cloud models have a bigger budget than small local models, but it's still limited.

We spent quite a bit of time crafting a very short, efficient personality description that gets good results while using almost no tokens. That leaves as many tokens as possible for your actual question, the game data, and the command you want handled.

If we let people freely edit/add personalities or long custom phrasings, it's very easy to blow past the token limit, and/or introduce LLM confusion, especially on local models. When that happens the AI starts hallucinating, lying, forgetting things, giving weird answers, or just breaking completely.

Right now we offer four carefully tuned styles that stay lightweight:
• Professional (think a military pro)
• Casual (a co-worker, still professional, but not strict as military)
• Friendly (like an old friend, adventure partner)
• Unhinged (your drinking buddy, but keeps the profanity in check)
• Rogue (a gangster rogue AI, may use profanity if model allows.)

Say: "Change personality to <Casual/Friendly/Rogue>" and app will swap it. Same goes for cadence (Imperial, Federation or Alliance). Trese only work with cloud LLMs, and Google STT at the moment.

NOTE behavior will depend on LLM. For example Grok can do Unhinged and Rogue much more convincing than ChatGPT.

That gives variety without risking the balance.

Quick note: these personalities only really shine with cloud LLMs + Google TTS. On local models (and current local TTS) personality stuff is disabled because small models can't really "act" convincingly or at all, and it would just waste tokens for little no gain.

So for the foreseeable future we're keeping the core personality instructions non-editable to protect reliability, especially for people running local small LLMs (8b) on weaker hardware.

If there's a specific tone/behavior you really miss, feel free to describe it in this ticket, we can consider adding it as a fifth built-in style if it can be done efficiently.

Capabilities

For the commands and capabilities, you can ask the app yourself. For example “How can you help me with ?” and it will give you instructions on how to use it. “How can you help me with trade?” LLM will understand you want help with trade and will provide explanation how that works.

For the commands, it can operate ship and SRV controls but it can’t operate weapons or steer the vehicle.

If you want to see a complete list of commands the truth is in the code:

Commands

Queries

You do not have to memorize them, or utter them precisely as listed in these files. Those are hints for LLM to infer your whishes. LLM will understand “request docking” same way as it will understand “contact tower and request parking spot”. That is the beauty of it. No need to memorize the commands, only understand the concepts.