The open-source Deepseek-R1 (available in the API as deepseek-reasoner) has been making the rounds as a model that's cheap, smart, interesting and just a little bit deranged. This prompt aims to influence the generated reasoning to put a greater focus on general coherence of events and writing.

Requirements

The prompt now requires the NoAss extension to turn the entire context into a single user message.

1. In ST go to Extensions > Install extension,
2. Put in the URL https://gitgud.io/Monblant/noass and install it for yourself or all users (doesn't matter which if you dont use ST's multi-user functionality),
3. Configure it like this.

Why?

The idea is to send the entire context as a single User message, avoiding the User / Assistant alternating which is tuned for problem solving rather than roleplay.

Setting the first block to User role makes ST turn every block below it into User as well (check the console!) This has no effect on chat history, though; NoAss is used to turn Assistant messages in chat history into User ones too.

Usage notes

• This is tested on the official Deepseek API. There are some differences between it and unofficial hosting of the same model on OpenRouter, Kluster and others. I can't accomodate for them without making accounts everywhere myself, but if you discover an issue, please let me know.
• Update to SillyTavern Staging for the Show model thoughts checkbox when using R1 via official API. Currently not available in the source Custom (OpenAI compatible), but thoughts are visible in console if you disable Streaming.
• There's no prefill in use at the moment.
• The official API ignores all sampler parameters (Temperature, Top-P etc.) Some other providers hosting the same model have temperature available.
• Have a look at the rundown (NSFW) for more details on the model.

Changelog & old versions