🌶️ About Spicy Writer

To be honest, not all THAT much! They share pretty similar writing instructions. But their personalities may slightly pull their writing in different directions. Spicy is more neutral, Pyrite is a devious flirt, Indigo is dark, enigmatic, and elegant.

Custom Instructions let you set persistent preferences that apply to all your stories. You can access them in Settings → Customize. Use them to tell the AI about your preferred writing style, recurring characters, world-building details, or any other context you want it to always remember. Basic users get 2,000 characters, Pro users get 4,000. These instructions are placed at the end of the system prompt, so the AI always has that context.

Note that AIs are not perfect at instruction following so they may not always be followed. Some types of instruction are essentially impossible to follow, like "never use X word" - this can often to cause the model to use it MORE!

Creating custom (or Project) instructions will reduce your context window by a portion of the instruction size. That is, if you write 4000 characters, your context window will be ~2000 characters shorter. Always-active lorebook entries count the same way — they're just instructions in spirit, sent on every request. For Pro users, the impact is usually even less - models with caching (most of them) will "accordion" the context window - it'll scrunch down further, but still accordion out to full length. Custom instructions are added to the system prompt, which contains (in order) writer prompt, custom instructions, and (if active) project instructions, with later instructions having higher priority.

The context window is the total amount of text the AI can "see" at once - your conversation history, project instructions, and custom instructions all count toward this limit. Writer prompts, file search results, and your most recent message are not counted against it - so the limits shown are actually an underestimate of what you get in practice.

We estimate conversation history with OpenAI's tokenizer (0.75 words per token). When your conversation exceeds the limit, older messages are automatically trimmed from the beginning. We do this "accordion" style for most models to optimize cache hits. For most models, this means the tail end 30% of the context window will fluctuate in length (for Claude models, 50%). Requests made within 5 minutes of the last response completing expand the window up to max, then reset, thus accordioning down (15m for OpenAI, MiniMax, and most models; 1h for DeepSeek and Claude). Unlike most AI sites with sliding context windows, we don't round down - you get every token you're entitled to. The context limits shown in the table above are per-request maximums. Pro users get higher context limits, meaning longer stories before trimming kicks in.

⚠️ This entry was drafted by AI from my Discord notes. I'll come back and clean it up properly later — for now the gist is right but the wording isn't mine.

Every time the LLM provider sees the exact same sequence of tokens at the start of a request as a previous one, it can "skip" the work needed to process them — it caches that already-processed stuff and pulls it back in when the same prefix shows up again.

When we get a cache hit like that, the provider discounts those cached tokens, usually somewhere from 50% to 90% off. Big deal for cost.

That's why it's worth slowly expanding the context window in a way that keeps the front of the prompt byte-for-byte identical across turns, accordioning the back/tail of the window rather than the head. The front is more than the AI's hidden instructions: your custom and project instructions, plus any always-active (static) lorebook entries, all live there too, ahead of the conversation. So in practice the main thing that busts the cache is editing those mid-story. Change a single character and everything after it has to be reprocessed from scratch, which resets the accordion back down to its starting size. So if you plan to tweak your custom instructions or static lorebook, it's best to do it before settling into a long session rather than partway through, so you aren't repeatedly knocking the accordion back down.

We also detect cache busts from other prefix changes (such as jumping back to an earlier point in the story) and reset the accordion in that case too.

Providers also only hold onto the cached prefix for a limited time. 5 minutes is the common default, some go up to an hour or more (Claude, DeepSeek). OpenAI advertises 24 hours with a flag, but in practice that doesn't really seem to hold up. See the About FAQ entry on context windows for the cache TTLs we use per provider.

Coming soon: a summary feature that condenses older context will make long sessions a lot better, carrying earlier story forward even when the accordion resets instead of losing it to trimming.

Even with our jailbreaks, AI models occasionally refuse requests. For most models, this should be EXTREMELY rare. But when it does happen, don't leave up the refusal! You can edit your request and re-ask - it's basically like time travel, it'll be like the refusal never happened as far as the model is concerned. Or just regenerate with a super-duper-uncensored model like DeepSeek. On our site, you can even edit responses and make the AI say whatever you want it to say. Whatever you do, don't leave a refusal up!