TL;DR
When you add requirements to a story as the conversation goes along (first the genre, five messages later a character, then you suddenly remember a special phrase that has to be in the text), the final story comes out structurally worse than if you'd given all the requirements at once in a single prompt. What's more, the model can formally tick off every item on your requirements list and the text will still fall apart on pacing, cause and effect, and setting up the payoff.
The thing is, the model can't go back and rewrite already-generated text in light of new information. It writes linearly and just "stitches" the new requirement onto what's already written, as if inserting a paragraph into a finished text rather than rebuilding the story from scratch. The spread across models is huge: one loses just 12% of quality under incremental delivery, another more than half.
The researchers introduced a metric called Creative Integrity, a joint score of "did it meet the requirements" + "did it produce a coherent story." With incremental delivery, models on average retain only 71% of the integrity they achieve with a full brief up front. And even when two versions of a story formally meet the requirements equally well, the version written "with the full brief as input" still comes out structurally more cohesive.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint (what happens and what to do)
WHAT BREAKS QUALITY:
Requirement 1 → the model writes a chunk of text
Requirement 2 (3 messages later) → the model "stitches" it onto the finished text
Requirement 3 → another patch, not a rebuild
= the structure crumbles, even if every requirement is formally met
WHAT WORKS BETTER:
All requirements at once in one prompt → the model plans the whole story → higher coherence
OR, if requirements came up along the way:
Collect the accumulated list → ask for the whole text to be rewritten from scratch
Example
Task: You're writing a script for an ad sketch about a delivery startup for Reels/TikTok. You start a conversation with ChatGPT and remember details along the way: "let's have a meme about the courier," "and at the end, a twist that the customer used to be a courier himself."
How people usually do it (bad):
Prompt 1: Write a sketch about food delivery, 30 seconds, funny.
[the model writes]
Prompt 2: Add a joke about the courier always being late.
[the model adds the joke to the finished text]
Prompt 3: And make a twist at the end — the customer turns out to be a former courier.
[the model tacks the twist onto the end]
The result is a sketch where the lateness joke sticks out as a separate chunk and the twist isn't set up in advance: the viewer won't get why it's surprising and funny.
How to do it better:
Prompt (everything at once):
Write a script for an ad sketch about food delivery for Reels, 30 seconds.
It should be funny. Must include:
- a joke about the courier always being late
- a twist at the end: the customer turns out to be a former courier
Build the plot so the twist is set up in advance —
the viewer should see the beginning in a new light once they learn the truth.
Or, if the requirements already came up during the conversation:
Here are all the requirements for the sketch I've given you in this conversation:
[list]
Rewrite the script from scratch, in full — don't edit the previous text,
create a new version where all these requirements are built in from the very start.
Result: The model plants "hooks" for the twist in advance and weaves the joke into the overall rhythm of the scene instead of inserting it as a standalone line. The final text comes out more cohesive: the reader or viewer won't feel the "seams" between the added pieces.
Why it works
The model generates text linearly, word by word, and when you add a requirement after part of the text is already written, it has no mechanism to "go back and rewrite everything with the new condition in mind"; it just slots the requirement into what's already there. That's the weakness: a local fix instead of global replanning.
The LLM's strength is that it's excellent at planning the whole story if it sees all the conditions at once before it starts writing. That's exactly why a complete brief up front works better: the model can think through the story arc in advance, set up the twists, and distribute the pacing.
Control lever: if you feel your task has piled up a lot of edits over the conversation, don't ask to "add one more thing"; collect the final list of requirements and ask for the text to be written from scratch. That's the only way to make the model replan instead of patch.
Limitations
⚠️ Only open-weight models were tested, not top closed models. It's unknown how strongly the effect shows up in the latest GPT-5/Claude/Gemini; they may hold structure better under incremental delivery.
⚠️ The effect is strong specifically in creative writing with 5–9 dialogue turns. For business texts, where only factual accuracy matters rather than narrative coherence, the degradation may be smaller or of a different nature.
⚠️ Automated evaluation doesn't always match human judgment, especially on originality: the automated judge thought incremental delivery even slightly improved originality, while human readers thought the opposite. Don't treat automated scores of creative text as absolute truth.
How it was studied
The team took 160 short-story assignments in six genres (fantasy, mystery, romance, science fiction, historical fiction, comedy) and made two versions of each: the full brief at once and the same brief split across 5–9 dialogue messages. They ran them through six open models, yielding 960 paired comparisons.
Quality was assessed with a checklist of met requirements and five text parameters (structure, writing craft, originality, genre fit, characters). Some of the scoring was done by an LLM judge and some by three human raters on 50 pairs, to check whether the automated scores matched human ones.
The most interesting part: even when they kept only the pairs where both versions met the requirements checklist equally well, the incremental version still turned out structurally weaker. This means that checking "were all items met" isn't enough: a story can formally pass every checkbox and still be worse assembled as a story. The spread between models (from 45% to 88% of quality retained) shows that resilience to incremental edits is a separate property of a model, not directly tied to its overall capability.
Resources
The Effects of Incremental Instruction Delivery on Language-Model Creative Writing — Anshuman Singh, Haseeb Yaqoob, Abrar Eyasir, John Manavalan. Dataset and code: github.com/solusops/SISTER-2026-Team19