Feature Request: Persistent, Editable Style & Behaviour Controls for Consistent Creative Workflows
Context & Use Case
As a fairly new, but already devoted creative user of ChatGPT, I’ve been exploring the boundaries of storytelling, character development, and scene refinement through GPT-4’s capabilities. My experience has been most productive when the model is able to adapt to my preferred style: slow-burn progression, atmospheric depth, emotional nuance, and richly developed character psychology. When it works, it’s extraordinary. However, maintaining this consistency has proven nearly impossible, especially across longer projects.
The Problem: Behavioural Drift & Default Regression
Despite early sessions starting strong, there is a frustrating, repeating pattern:
- The model initially adapts to my instructions.
- It produces well-structured, immersive scenes in line with my tone.
- After about ten to twenty replies/comments it begins to default back to its typical style: shorter sentences, fast pacing, script-style formatting, and tonal simplification.
- Any attempts on my part to redirect or reinforce my preferences become less effective, and I’m forced to either rewrite large portions myself or restart a fresh chat (which resets all nuance).
- This creates significant fatigue, blocks creative momentum, and in some cases has caused me to abandon stories altogether, at least for now.
My request is for a more functional AI, a model that can hold onto its adapted state and not regress without cause.
What I Need (and Propose):
- Editable Persistent Memory:
- I should be able to tag a chat or writing style as the “baseline” for a story or character, and update or tweak that memory directly.
- I should also be able to lock in stylistic preferences (e.g., avoid script-formatting, maintain paragraph structure, prioritise immersive prose).
- Behaviour Toggles:
- Allow me to toggle default writing behaviour. For example:
- Avoid script-style formatting
- Maintain slow pacing / descriptive paragraphing
- Use immersive, character-internal perspective
- Use UK-English instead of US-English (preference)
- Cross-Chat Continuity:
- GPTs offer surface-level customisation but don’t maintain meaningful memory across sessions unless hardcoded, and even then, they are still prone to drift.
- A true ‘project mode’ is needed — where a model remembers voice, structure, and story continuity.
Without these options the gradual drift away from the tone and structure I explicitly ask for is not only frustrating; it actively undermines collaborative story progress and destroys creative momentum. Eventually, I have to abandon the chat and start over. This process is emotionally draining, creatively discouraging, and completely avoidable if more persistent customization tools were in place.
Real Examples: GPT-4o vs GPT-4.5
Here’s a side-by-side comparison from within the same chat, based on the exact same prompt, with the only difference being that I switch from GPT-4o to GPT-4.5 after copy-pasting said prompt.
GPT-4.0 output
The mess still lay there. Bits of broken wood. Shattered ceramic. The weapon. The sweat-soaked shirt from before. It wasn’t just chaos—it was dangerous. Sharp edges. Splinters. Blood on the floor, dried into thin brown lines. Nothing cleaned. Nothing moved.
And X hadn’t sent anyone to fix it.
Y stared at it for a long time.
Then he crouched down and picked up the largest piece of ceramic near the bed—white, jagged, glinting at the corner like it could still bite him if he stepped on it wrong. He wrapped it in a scrap of torn cloth and set it aside. Another fragment followed. Then another. The motion was easy. Methodical.
It wasn’t about obedience.
It was about control. About making the space usable again. Navigable. Safe.
Result: staccato sentences, shallow tone, reads almost like a script
GPT-4.5 output
When Y returned to the main room, the mess he’d left was still waiting. Shards of ceramic and splintered wood littered the floor near the bed, sharp edges catching the dim overhead lights. He stared at it for a moment, then slowly moved forward, crouching down to pick up a jagged piece of ceramic. He wrapped it carefully in a torn scrap of fabric, setting it aside so it wouldn’t slice into his feet if he stumbled past in the dark.
It was practical. Necessary. Just making the room safe again, manageable—
Making his space safe.
The thought hit him harder than any physical blow. He froze instantly, breath catching in his throat, his stomach twisting sharply at the idea. He hadn’t meant to think it, hadn’t meant for it to slip past the guard he kept around himself, but there it was, real and ugly. His space. He was already treating this place like it belonged to him—like he belonged to it.
Y stood abruptly, backing away from the mess as though it had burned him, palms sweaty, pulse pounding in his ears. He retreated until his legs hit the bed and sat down heavily, gripping the edge of the mattress tightly enough for his knuckles to ache.
He didn’t move again, just sat staring blankly ahead, letting the quiet of the room wrap around him, feeling every second pass as the word echoed inside him like a taunt.
Result: drastically better flow, pacing, depth, atmosphere
While there is nothing inherently wrong with the first version, it is stylistically empty. It lacks the immersive pacing, internal monologue, and emotional nuance that makes the second version more compelling and matches my writing style. In the first version you can clearly notice how the GPT reverts to short sentences, paragraph gaps, and summary-like narration.
Why Custom GPTs Aren’t Enough
I’ve tried using Custom GPTs that have been specifically created or designed to help with (creative) writing. I have given each of these instructions tailored to my preferences with regards to my tone of voice and style. While this helps initially, the same slow drift as with GPT-4o happens over time. Instructions about not using short sentences or a movie script-like style are either ‘forgotten’ or ‘ignored’ due to ChatGPT’s internal framework forcefully overriding any instructions concerning style, whether these are given in-chat OR have been written into the ‘core memory’ as instructed. Without a more robust memory system—or a way to ‘recalibrate’ style mid-session—Custom GPTs don’t solve the underlying problem.
Why This Matters
While there is nothing inherently wrong with the first text – it’s functional and does convey the scene – it is not the style I use for my stories. For users like me who are heavily invested in long-form creative projects, GPT-4.5 represents a near-ideal writing partner. It respects atmosphere, knows how to build tension, and writes in flowing, immersive prose without needing the kind of constant – and I’ll be honest insanely frustrating – corrections.
Without a way to keep the tone I want consistent, that ‘magic; degrades quickly, annoyingly, and to the point where I feel like I’m banging my head against a wall and are on the verge of just abandoning the project because GPT4o is incapable of properly adhering to my instructions due to its far too dominant default writing style. If OpenAI gave me the tools to lock, edit, and preserve the version of GPT that works for creative users like me, it would be transformative.
Access Limitations Compound the Problem
To make matters worse, GPT-4.5 — the version that consistently hits the mark for immersive storytelling, paragraph structure, and psychological nuance — is currently only available in a limited and inconsistent fashion. While Plus users do have access to GPT-4.5, that access is constrained by an undisclosed cap on replies that seems to vary wildly between individuals. No one — including those actively discussing the issue on the OpenAI Developer Forum — has been able to determine what governs this limit or how long access lasts before reverting to GPT-4.0 as far as I can tell.
This inconsistency makes it impossible to rely on GPT-4.5 for long-form projects. The model that delivers exactly what’s needed becomes a temporary tool, revoked without warning, right in the middle of crucial narrative work. It undermines planning, continuity, and confidence in the platform, and forces users to fall back on GPT-4.0.
Which brings me to the other issue: GPT-4.5 is also locked behind the much more expensive Pro plan at $200/month — an option that’s completely out of reach for many independent creatives like me. The difference between GPT-4.0 and GPT-4.5 is massive: the former inevitably slides into generic output no matter how carefully I prime it, while the latter delivers the tone, pacing, and complexity I need without constant correction.
In short, the version that works is either restricted by unclear, arbitrary limits or locked behind a steep paywall — neither of which serves the needs of writers doing serious, sustained creative work.
Final Note
I want to make it absolutely clear that this isn’t a complaint from someone who anthropomorphizes the model. I know what GPT is, and what it isn’t. What I’m asking for is a tool that works consistently and adaptively. One that doesn’t fall apart halfway through a carefully structured story. One that remembers how I write, how I think, and how to work with me—not against me.
I hope this feedback can be taken seriously. I’m more than willing to elaborate, collaborate, or help test features if that’s ever an option. Thank you.