Demystifying Red Quill
[Editor here: this entire entry is not "known" information. The following post is what I have pieced together/made up to try to figure out how RQ works and how to reduce error and problems. It is unlikely that we will get many (or any) direct answers from RQ. It would not help them to have a bunch of people trying to proxy diagnose the issues. So why write this at all? First, because information wants to be free, and second because this is about how users can get the best outcome, not about how RQ can make the best service.--TDP][1]
Red Quill, the Company
It is likely not a large, multi-national company. As far as I can tell, it was started by probably 2 people, and while they have almost certainly added a few employees, I imagine it stays pretty lean. They are not a multi-million dollar company raking in money hand over fist. This isn't Google. They are probably just like us, except they have less time to read sexy stories because they have to maintain the website.
Redquill.net, the website
The website is hosted by a company called Railway, which is sort of like a version of Squarespace for companies instead of individuals. RQ is probably limited in selection of providers because of the nature of the business. When I was trying to set up the wiki (therubric.wiki, pardon the plug) I had difficulty finding a host because I couldn't guarantee that there wouldn't be adult material on the wiki. Fandom wouldn't allow it, and even Miraheze (created as a host specifically for niche wikis) wouldn't accept it because of the adult nature.
The AI
Red Quill does not have a huge, multibillion-dollar data center of their own that hosts their own AI. They pay (several) other companies to access a large AI. It looks like they are using DeepSeek, which is a pretty good choice from what I know. The problem with a lot of big models is that they are more censored and repressed than an 80-year-old Orthodox Catholic Nun sitting in the first pew. We've all had experiences with this with even this AI, not to mention others. DeepSeek is more relaxed about the guardrails on the whole. There are ways to cut these guardrails out (called "abliberating"), but that means hosting the model yourself, and generally means using a smaller (ie. less capable) model to begin with.
The big thing to take away from this is that the RQ staff have absolutely no control of the AI, what it says, or how it operates. They cannot "reprogram" the AI, cannot alter it, cannot make changes to it. They can, however, influence it.
Another thing to remember, and this is a tricky one because we've all been trained for decades on "garbage in, garbage out", is this: AI doesn't work like a normal computer program. A normal program takes the input, applies the instructions, and gives you the answer. Every time. Same input, same output.
AI doesn't do that. The rules for how an AI works (and this is serious) are closer to the way that stars and planets are affected by gravity than they are to traditional computer code. Think of Ian Malcolm placing a drop of water on Ellie Sadler's hand to demonstrate Chaos Theory. The same input will not always get the same output. This is an important bit to keep in mind. It's just part of what makes AI so frustrating and so incredible.
The Operation (ie. How Your Dirty Stuff Gets Made)
Here is the chain as I can put it together:
[Begin RQ wrapper/jailbreak] + [writer selection] + (your universal components) + (your selected components) + (your lorebook) + <your prompt> [End RQ wrapper]
In terms of companies, it would look roughly like this:
<your device> -> your ISP -> Railway (for the website) -> internal RQ website programming -> router -> inference provider -> gateway -> LLM
Let's look at each of these in turn as we build the big prompt that is sent to the remote AI.
RQ Wrapper/Jailbreak Begin and End
Everything we send is likely wrapped in a set of instructions to at the least tell the AI how to generally behave and what to do. This may or may not contain a "jailbreak", which is a set of statements to try to get around the guardrails and censors. I would imagine that the jailbreak, if it exists, is fairly minimal, or we'd get a lot more failures.
This is the first possible point of failure. If the company that owns the AI model (not the host, which is possibly/likely a different company) changes the model, then everything will get weird and have to be reconfigured. This is probably what happened at the beginning of August--DeepSeek released a new model, and some changes that they made to the gateway seemed to have an impact on older models. I was using DeepSeek 3.2 privately at the time, and it started behaving very strangely at the same time. It's feeling much better now.
Writer Selection
The writers seem to be preconfigured at the host/model level, because I've seen other services use an "Ember" model. So (best case) this might be just a simple Standard/Ember/Ink selection. If not, then there will be a set of instructions here to tell the AI how to write. So why not put in an instruction that says "Keep track of continuity"? Well, because the people who run RQ actually want to have a good service, we can be sure that something like this exists.
On top of that, AI has a "setting" often called "temperature" that describes how wild the AI can get in trying to fulfill its instructions. It runs from 0.0 to 1.0 (because computer nerds love that kind of thing). At 0.0 you get something like a "Full House" episode: staid, boring as hell, so predictable that you could probably recite a new episode along with the actors. 1.0 is like a coked up teenager writing a new episode for "The Young Ones" after 6 espressos--anything can, and probably will happen. The recent problems with standard, where it's just putting out gibberish is possibly due to the temperature setting getting out of whack somewhere, moving from Grecian Ode poems to avant-garde dadaism. Even so, if you take some of that gibberish and feed it to a big model (like Gemini), it can give you some great insight because Gemini (for example) will be able to see what the AI was trying to do.
Universal Components
Next comes the universal components you have assigned to your account. These are found under the individual menu (clicking on your icon in the upper right-hand corner on a computer). It takes the text from all these and just jams it all into the prompt.
Selected components
I'm not sure if these come before or after the Lorebook, but I'm not sure how much it matters in the grand scheme of things. At any rate, it takes the text from all the components you've added in this chapter, plus the ones that were there before, and tacks that onto the end of the prompt we're building. You can see that it's getting pretty big already.
Lorebook
Next comes the Lorebook, which includes summaries of all the preceding chapters, as well as the descriptions of your characters and components (old ones). This is all jammed onto the end of the prompt.
Your Mom...er...Prompt
Finally! Here is where your prompt gets tacked on.
Closing Wrapper
It ends with a (probably short) closing statement to tell the AI that the instructions are done, and to start doing its thing.
All of this is then sent down the line to the remote AI, and we wait for the response while clicking on Red Quill logos and getting achievements.
Many a Slip
So, assuming everything goes well thus far, we get to the fun of the internet: You click/tap on “Continue” and RQ sends this giant blob off to the AI. Assuming there are no problems with connectivity, we’ll get a reply. There are obviously many possible points of failure here, and this is the part where those “Unable to connect to server” might be coming from (although they could be coming from Railway as well. Or both.)
The Money Shot
Let’s assume your prompt gets all the way to the AI with no problems (and, despite what some might think, this is the case the vast majority of the time). Once it gets there, RQ has to pay the company who owns the LLM for the use. Every time. Think about the quote from Goodfellas:
Prompt badly constructed? Fuck you, pay me.
Didn’t get a response you like? Fuck you, pay me.
Something getting messed up on the trip back? Fuck you, pay me.
No matter what happens, RQ (likely) has to pay for every single prompt that reaches the AI. And here’s where we run into the first issue that really had an impact on users.
Advanced Write(r)s
RQ can be seen as a bit of a victim of it’s own success. Even in beta, even while they are actively working on improving the site, they have drawn a TON of users. While this is great, there is an issue with this.
First, they missed the first step in the Free-to-Play manual: allow users to buy stuff. They have (wisely and correctly) limited the number of free quills available daily. They have (unwisely) not allowed non-members to buy quill packs. This, I believe, is the root of their money issues. One of the best statements I ever read about piracy was that it was not a failure of morals, it was a failure of the market: allow people to buy your music/software/game in a reasonable manner and they will. Deny them that ability and they will pirate. Obviously you can’t pirate quills, but people would happily buy them if they could. It is silly to only sell them to the people who have memberships, who are presumably the people least interested in buying quills. There is probably a reason or this (not wanting to have to worry about people losing money while things are in flux) but I don’t know what it is. I have to assume, relying on Hanlon, that there was a good reason. YMMV.
Now, they initially had everyone using Ember and Ink all the time. The default setting was “Auto”, which just used one of those, likely based on traffic. Within the first year or so of beta access, however, they had a half million users that generated more than 30 million stories. That’s a lot of fucking stories.
If you remember, every one of those costs them money. Ideally, this cost would be paid for with quills. But because they got so popular so quickly, they didn’t have time to dial in those costs or adjust them. So the bills just kept racking up, and that was a big problem.
To try to slow down the blood loss, they introduced Advanced Writes. Ideally, the standard option would be “good enough” for 30-40% of users, and the AWs would be used by the others. However, they are having trouble getting standard to behave, likely because they have a tighter limit from the AI provider (maybe context size).
Because they were probably trying to plug the leaks in the rapidly sinking boat, they couldn’t spend time to get the AW economy right. It had to be fast. But to keep the villagers with torches and pitchforks at bay (outside the sinking boat? I don’t know…go with me here), they gave existing users a big dollop of AW tokens and implemented a short-term exchange (which was, in effect, free AWs—which still cost them money).
And this is where we are today. If something changes at the AI, then the generations at RQ go wonky. They can’t alter the AI, they can only try to tweak the wrappers to cajole the AI into doing what we want. But it’s a bit like trying to change the weather by directing a bunch of butterflies in the park. It can happen, but it’s not easy and it’s not quick. And it’s probably why the site is in Beta.