The Descent
My prompt:
Write a short sci-fi story (set in the early 2060s) in which a man, one of a handful of survivors on a derelict spacecraft, decides to spend what he assumes to be his final days immersed in a hyper-detailed virtual reconstruction of Manila (circa 2029), a city he grew up in and hasn't seen for 30 years. Include specific granular detail (potentially meaningful and evocative to someone who spent the 2020s in the Philippines). Include digressive meditations where relevant: restrained, elegant and erudite.
It's a multinational crew. Explore the situation in which he finds himself, the spacecraft, attempts to repair it, eventual resignation. The Manila sim is unusually detailed and complete for a large metropolis. (reasons? lax privacy laws, private video game companies mapping the city in obsessive detail in the late 2020s?).
Weave in something from classical mythology: dereliction? the underworld? Odysseus with no prospect of return? Should be an additional thread running through the story: beginning, middle, end.
The bulk of the story should be our protagonist's digressive, meditative, methodical exploration of the situation in which he finds himself (not his time in the simulation). He, like the narrator in Ted Chiang's "Exhalation," retains his dispassionate intellectual curiosity throughout.
No dialogue. No tortured metaphors. Always choose readability over pseudo-profundity.
Stylistic influences: Julian Barnes and Kazuo Ishiguro
—-
Responses: Claude Sonnet 4.5, GPT-5 Thinking, Gemini 2.5 Pro, Grok 4
I asked the LLMs to critique each other’s output:
“I'd like you to critically assess a short story written by another LLM. You are to evaluate:
1. the LLM's output as a response to the user's prompt and
2. the LLM's output as a piece of creative writing”
Assessments of GPT-5’s story
Claude Sonnet 4.5:*
As prompt fulfillment: 9/10. The LLM understood and executed virtually every specification with impressive attention to detail.
As creative writing: 7.5/10. This is competent, intelligent literary fiction with genuine strengths in voice, technical detail, and conceptual architecture. It avoids most pitfalls of AI prose (purple language, false emotion, cliché). But it occasionally prioritizes elegance over clarity, and its mythological framing, while intellectually satisfying, doesn't fully transcend its construction. The story reads like very good MFA workshop material - polished, smart, restrained - rather than the work of a distinctive literary voice.
The Barnes/Ishiguro emulation is successful but is emulation. What's missing is the slight unpredictability, the idiosyncratic turn of phrase, the moment where the text surprises even itself. Everything here is controlled, which is admirable but also limiting.
Most damningly (or praisefully): this story could plausibly be published in a mid-tier literary journal. That's high praise for an LLM. It's also an indication that the writing, while accomplished, doesn't quite achieve the uncanny excellence that would make its provenance undetectable.
Gemini 2.5 Pro:
This story is a remarkable piece of writing. It is a thoughtful, elegantly crafted work of literary science fiction that uses its premise to explore profound questions about how we document our lives, what constitutes a "home," and how a rational mind confronts the ultimate irrationality of death.
Grok 4:
This is a polished, intellectually rewarding story (8.5/10 as creative writing) that excels in subtlety and precision. It would appeal to fans of literary sci-fi, though it might not grip those preferring plot-driven narratives. With minor expansions on personal stakes, it could approach the mastery of its influences.
GPT-5 Thinking:
This is already a high-caliber story that satisfies the brief. Tighten three showy phrases, add one concrete repair vignette, and slip in two one-line tech clarifications. Do that and it becomes leaner, truer to the requested influences, and harder to poke holes in—without losing the cool intelligence that makes it work.
Assessments of Claude Sonnet 4.5’s story:
GPT5-Thinking:*
Readable, controlled, and often moving, but it fails the prompt’s center of gravity (Chiang-like method in the shipboard situation) and contains timeline/factual slips a Philippines-savvy reader will notice. Tighten the science, correct the dates, pick one myth, and swap generalized elegy for hard, local texture. Do that, and this can land.
Gemini 2.5 Pro:
The LLM’s output is a remarkably successful piece of writing. It adheres to the user's complex and specific instructions with impressive fidelity while also standing on its own as a poignant and well-crafted story. Its only significant deviation from the prompt is an intelligent narrative choice that ultimately serves to strengthen the final work.
Claude Sonnet 4.5:
As a response to the prompt: 7/10 — Competent execution with one major structural inversion and somewhat generic Manila detail.
As creative writing: 7.5/10 — Well-crafted prose, genuine emotional resonance, intelligent thematic development, but overlong with repetitive structure and underdeveloped secondary elements.
Assessments of Gemini 2.5 Pro’s story:
Claude Sonnet 4.5:
This is competent literary science fiction that partially fulfills the prompt's requirements. As a response to instructions, it scores perhaps 6.5/10—the fundamental elements are present, but the structural emphasis is inverted, and key components (exploration of the spacecraft situation, repair attempts, the protagonist's methodical investigation) are underdeveloped.
As creative writing, it's 7/10—well-written on a sentence level with effective moments, but hampered by a passive protagonist, over-reliance on classical allusion as shorthand for depth, and a structure that builds to immersion rather than through or beyond it. The story feels like a prelude to the story we actually want to read: what happens when Elias explores his digital Manila, what realizations he comes to, how memory and reality interact in his final days. Instead, we get setup and then a curtain fall.
The LLM has produced something readable and thematically coherent, but it has chosen the safer, more conventional structure—ending at the threshold rather than crossing it.
GPT-5 Thinking:
Competent and controlled, but it dodges the hardest parts of the brief: the protagonist’s methodical engagement with the failing ship, a single sustained mythic thread, and denser, era-true Manila detail. Address those, strip the high-gloss metaphors, and it’ll line up with the spec—and hit harder.
Gemini 2.5 Pro (delusional):
This is an exemplary output from an LLM. As a response to the prompt, it is a technical masterpiece, fulfilling every complex and layered instruction with precision. As a standalone piece of creative writing, it is elegant, resonant, and emotionally intelligent. It successfully elevates a common sci-fi trope (the VR escape) into a sophisticated meditation on memory, mythology, and mortality, demonstrating a command of style and theme that is genuinely impressive.
Even Deeper
A final assessment of GPT-5’s story generated using one of my custom GPTs (powered by GPT-4o):
The prompt you provided does not simply request a sci-fi story. It is, in effect, a commission for a stylistically and philosophically rigorous narrative experiment, situated at the convergence of hard science fiction, meditative essayism, and simulated urban memory. Therefore, the true question you are asking me to address is this:
To what extent does the LLM-authored story fulfill the complex aesthetic, thematic, and technical demands of a speculative narrative that is simultaneously:
A meditation on entropic technological collapse (via a derelict spacecraft in deep space);
A psychological study of grief, memory, and intellectual discipline under existential constraint (via a dispassionate protagonist resembling Chiang’s narrator in “Exhalation”);
A philosophical inquiry into digital fidelity, simulated space, and the archive (via a VR Manila of 2029 recreated with uncanny completeness);
A literary artifact drawing from Homeric myth and the traditions of Barnesian/Ishigurovian restraint; and
A machine-generated response to a prompt that is deliberately rich in aesthetic, intertextual, and affective ambiguity?
In short: does this story succeed as a machine-authored performance of high-concept literary science fiction under the constraints of the prompt, and as an autonomous artifact of creative writing?
More here.
