The leech guide to LLMs: How to coombot for free

Back to Main Page

Preamble

The goal of this rentry is to give you a headstart with chatbotting. I'm assuming you know how to use Git and are able to google stuff on your own.
Consider these to be hints moreso than a full fledged A-Z guide.
System Requirements: Toaster. Pretty sure you can set this stuff up on phones too.
Useful links:
4chan /g/ Look for the "AICG" threads.
Chub Site where people upload cards (these are the chat bots). Questionable quality on most of them.

Why use remote LLMs over Local?

Simply put, they are better. Once you try proprietary, anything local feels like ass.

Why can local image gen compete with proprietary, but for LLMs they don't?

Also a simple answer, the system requirements are much higher, and a lot more is behind closed doors.
Deepseek R1 is one of the best open-source models at the moment, but running it at full capacity and not a distilled version (basically a smaller one) would require 8 A100's. So unless you got 100.000€ to waste, that's not realistic. Meanwhile Deepseek R1 is also available for free through OpenRouter, which is a middleman between end-users and the various LLMs that exist. How can they offer it for free? Venture capital.
There ARE competent local models that fit into 12GB vram and all, but unless you have serious privacy concerns, you lose in every other aspect.
If you care about privacy then use a VPN. Realistically this doesn't do much. Just accept that John Thomas is gonna read your cookie coom slop.

Installing SillyTavern

SillyTavern is a frontend to do chatbottery.
First skim their Github page
Next you wanna clone the repo git clone https://github.com/SillyTavern/SillyTavern.git, then run UpdateAndStart.bat. If you used the webUI or Comfy before then a lot of this will feel familiar.
After installing, SillyTavern will open in your browser.

Connecting to a model

Step 1: Connecting Deepseek through OpenRouter.

• Click the Plug icon in the top bar (the second from the left).
• Change API to Chat Completion and the Chat Completion Source to OpenRouter
• It will prompt you to authorize it. Go to the near bottom of the window and click the authorize button. This will open the OpenRouter website and let's you create an API key. Follow the steps and it should throw you back to SillyTavern.
• For your model choose DeepSeek: R1 (Free).
• For your Model Providers choose chutes.
• Click connect.

Loading a Jailbreak

Jailbreaks (or JB for short) are a misleading term because they aren't solely used to circumvent censorship, for the most part these are just prompt presets to wrangle an LLM to do chatbotting better.
Once you are on Deepseek through OpenRouter, the top left button in SillyTavern opens the Chat Completion Presets. This panel let's you tweak your LLM on various parameters and also add additional instructions.
I recommend using a preset for now, you can always modify these after.
The preset I based my own prompt preset around is the AvaniJB. It's pretty popular and a couple of others are more or less based on it.

Step 2:

• Download the preset from the above link (Currently this is 2.5 Universal).
• To import the preset, open the Chat Completion Presets window, click the import preset button at the very top, then select the .json you downloaded.
• For DeepSeek R1 you want to lower the "Temperature" variable to around ~0.5. If the Output is ever PeNgU1N oF d00m tier or looks incoherent, lower the temp. The Temperature is a creativity variable, some models prefer a lower baseline. Think of it like CFG in imagegen.
• In the Preset panel, scroll down until you reach a bunch of Prompt panels. Notably they often contain an unfunny "did you actually read the readme check, if the output starts yelling at you to read the readme, maybe you should. The readme is usually the first prompt in the list.
• IMPORTANT: Importing a preset will reset your chosen model. You have to go back to the plug icon (API connections) and swap back to OpenRouter. If something is ever throwing errors, check if the API settings are still correct.

Optional: Regex. If a preset tells you to install some Regex, install them like this:

• Download the relevant Regex as mentioned. Currently the universal JB does not require one.
• To use the Regex, click the "3 cubes" Extensions button in the top bar, select Regex, click import. Once all the Regex are imported make sure they are actually toggled on.

Loading a chatbot

Click the top right button in the called Character Management. You will have some pre-installed chatbots. Click any of them.
Test if everything is working:
Step 3:

• Choose any chatbot
• Type "Hey bitch".
• Press Enter.
• ????
• Profit

You should be getting text in a legible form. If you don't, then go back and see if you missed something or I took a step for granted. Sorry! The most common error is the API connection (the openrouter stuff) resetting after importing a preset.
Next you may wanna download a chatbot from Chub, these are usually referred to as "cards".
Chub sucks. Features include: ESL, schizophrenic PDF files, and the profound stench of xenoestrogen.
Here is a Finnish Fox with a fat ass that is written decently. Click the "V2" Download button below the char's image. Once downloaded either Drag the image into SillyTavern, or Import her with a button.

Modifying characters is as simple as changing the text in their "card". For example, if you wanted to give her a fat cookie you would simply add that description, preferably near her other body descriptions.

Current good free models: Deepseek vs Gemini 2.5

IMO Gemini is better in most aspects, the 2 considerations over using DeepSeek over it are A: The rate limit, B: Privacy concerns (realistically both DeepSeek and Gemini could be 'spying' on users, I trust China more though).
If you want to try Gemini, change the chat completion source to Google AI Studio, Not the Gemini under OpenRouter, that ones censored like crazy. There you need to enter an API key, you can generate this by requesting it from google. Just search for "Gemini Api Key". Gemini can use the same Avani preset or use the specialized Gemini preset that's on his rentry. Either way you likely want to up the temperature a bunch, I've seen values from 1.1 to 1.8.
What I often do is start a chat on Gemini, then swap to DeepSeek when I run out of prompts. I think you get 25 daily?
You also have access to "Flash" versions of Gemini, these are dumber but still very usable and not nearly as rate limited.
And you also of course have the option to pay for an LLM. Claude used to be really good, but I haven't used it in ages.

How to write your own chatbot

TODO. Just gonna roughly outline how I write bots. LLM stuff has a lot less of a widely accepted correct baseline compared to image gen.
-Wrap sections into XML tags. LLM's can recognize and reference these fairly well. We actually have papers on this so it's not just schizo shit.
-{{char}} is a special syntax to refer to the character's name.
-Established characters often don't need as detailed of a description, given that the LLM can refer to known model knowledge. Examples would be Pokemon, those don't need excessive appearance descriptions but likely want a strong personality description.

<summary>
{{char}} is a 22 year old wolf who hails from the far north.
Yada yada, she was an orphan but she is super stronk yada yada.
</summary>
<Setting>
Medieval fantasy or some shit. There's magic.
</Setting>
<Appearance>
Big, strong, muscles, cookie.
</Appearance>
<Personality>
Cuddles.
</Personality>
<Fetishes>
Very very optional. Going overboard with sexual descriptions quickly turns the char into a coombot, especially on DeepSeek.
May actually be what you want.
</Fetishes>
<Roleplay Guidance>
1: Focus on slow-burn romance.
2: I use these as char specific rules.
3: Can basically add whatever you want.
</Roleplay Guidance>

My bots usually total around ~1000 tokens.

The first message is extremely important. It's the biggest decider to how characters talk, how the scenario is written, etc. I highly highly recommend writing this yourself and putting in some effort. You CAN prompt the AI to write your starting scenario for you by just doing something like:
[OOC: Write a compelling and romantic starting scenario to this roleplay between {{char}} and {{user}}. Focus on setting the scene in an adventurous and dark-fantasy writing style. Etc etc yada yada]. Then replace the initial message with the one generated. If you do this, make sure to at least replace the dialogue of your character with something more animated. Speech quirks are an easy way to guide a characters personality.

FAQ

What's a reverse proxy

A reverse proxy is a middleman server that sits between the LLM provider and you. OpenRouter is technically a reverse proxy, but in the chatbot community it usually refers to community hosted middlemen servers. These operate by crawling the net for API keys, and pooling them for the proxy users to use. Reverse Proxies have been declining, very easy to vandalize by just using the keys up maliciously. Most are closed communities now, and many are closing because crawling is getting harder. I saw one Reverse Proxy requiring users to grow their own mushroom colony over 2 months or some shit to prove they aren't bots.

What about AI chatbot sites that offer their own bots and LLM

These suck. This is like "effortgenning with bing" tier shit.