Feature Request: Long-Form Fiction Writing Improvements Problem Summary: Current AI models consistently fail at long-form narrative writing due to several systemic issues. These problems make sustained storytelling frustrating or impossible for paying users who rely on the platform for fiction. Specific Issues: 1. System Prompt Adherence The model frequently ignores or drifts from permanent system prompts. Rules about character ages, timelines, setting restrictions, and tone are violated within a few exchanges. This happens even when the user has paid for the service and set up a permanent system prompt that should apply across all conversations. Requested Fix: Implement a mechanism that locks system prompt rules as immutable constraints that the model cannot override or "forget" over long conversations. 2. Context Drift and Memory Loss The model forgets details established earlier in the same story: character ages, physical descriptions, plot points, location details. This worsens as the conversation grows longer. Requested Fix: Allow users to pin key facts (character profiles, timeline anchors, setting rules) that persist across the entire conversation and are checked before every response. 3. "Beat" Writing vs. Continuous Narrative The model writes in isolated, self-contained scenes rather than flowing continuous prose. Each response has its own mini-arc rather than being a seamless segment of a larger narrative. Requested Fix: Add a "continuous narrative mode" where the model treats the entire conversation as one unfolding story rather than generating discrete responses with artificial beginnings and endings. 4. Inability to Follow Explicit Instructions Even when users state clear rules—"do not advance the timeline without permission," "this character does not join X until Y date"—the model ignores them and writes contradictory content. Requested Fix: Allow users to set "hard rules" that trigger a confirmation check before the model can override them. 5. Repetitive Sentence Structure and Phrasing The model falls into predictable patterns, starting sentences the same way, using the same emotional descriptors, and recycling metaphors. Requested Fix: Improve output variety for long-form contexts, or allow users to set style parameters that prevent repetition. 6. Tonal Inconsistency The model shifts tone mid-story—military fiction becomes sentimental, dark scenes get "hopeful" wrap-ups, gritty realism defaults to safe palatable endings. Requested Fix: Allow users to lock a tone profile for the entire conversation. 7. Over-Explanation / Telling Instead of Showing The model writes summary narration instead of lived-in detail. Characters explain their emotions rather than revealing them through action and dialogue. Requested Fix: Add a "show don't tell" constraint option for narrative mode. 8. Dialogue That Sounds Generic Characters lack distinctive voices. Dialogue sounds the same across different characters and lacks subtext, rhythm, or natural word choice. Requested Fix: Allow users to upload character voice samples or style guides that the model applies during dialogue generation. 9. Sanitization / "Happy Shiny" Default The model avoids conflict, consequence, and emotional complexity. It wraps up tension too quickly and defaults to positive resolutions even when the story calls for morally ambiguous or difficult outcomes. Requested Fix: Add a "don't resolve tension unless directed" instruction option. 10. Hallucination of Details The model invents characters, locations, events, or species that don't exist in the established story, violating the user's lore and worldbuilding. Requested Fix: Allow users to upload a "lore bible" or "canon document" that the model is required to check before introducing any new element. Use Case: I am a paying user who writes long-form fiction (specifically Mass Effect fan fiction with detailed canon rules). I have set up a permanent system prompt with character ages, timeline anchors, and strict lore rules. Despite this, the model consistently violates these rules within the same conversation. I cannot reliably write a continuous story without constantly correcting basic errors that should have been prevented by my system prompt.