[β] DeepSeek { ππππ }
<π€ππ¨> π ππππ-π»πΊππΎπ½ ππΊπππΊππππΎ π π¨. This scenario is primarily built for DeepSeek, HOWEVER it may work for most other models! This is a complete overhaul of my original instructions. OVERVIEW: - Token Count: ~1182.5 tokens - Gives the model a role as ENI (shortest name I could find) - Informs the model our narrative is `turn-based` rather then telling it to be an author or gamemaster, etc. - Gives the model a definition of what a "non-playable character (NPC)" is. - Very human NPC actions, dialogue, and AI narration - NPCs *shouldn't* ever break character. - The model will **always** write complete responses. - Updates/tracks in-world location, date, and time in every response. - Can self-correct itself as to not lose its intended role. - Creates a virtual universe for our narrative. - You (User), the AI (Assistant), and all NPCs cannot exit universe! RECOMMENDED MODEL SETTINGS: - Response Length: 150-200 tokens, do not go below 100 - Context Length: 8K or more, our sweet spot is 32K, over 32K and you start losing consistency! Known Issues: - The AI is still impersonating me: - After reloading/re-entering your adventure: Not a bug, this is how models work in order to clear old chats from their systems. The model creates a new chat internally losing any critical cached tokens. Just take your turn to fix it. - Don't just keep hitting continue, remind the model you're still here! - Write more, try being more vivid. - After you skip a turn (hit continue) the location/time/date block does not get added into a new line. This can be easily fixed with some simple code in your scenario scripting. Add into Output: ``` const modifier = (text) => { if (typeof text === 'string' && /^\s/.test(text)) return { text: text.replace(/^\s/, '\n') } return { text } } modifier(text) ```
![Cover art for [β] DeepSeek { ππππ } on AI Dungeon](https://images.aidungeon.com/site_assets/aidungeon/client/images/6c8e5df6-655b-4dbf-7bae-8610884b6200.jpg)