The Exposition Problem That Breaks Hard Sci-Fi
Hard sci-fi has a tell. You can spot a manuscript written by someone who loves the genre but hasn't solved its central craft problem within the first three pages of any technical scene. Either the prose stops dead so a character can explain tokamak physics for four paragraphs, or the science gets so vague that the "hard" in hard sci-fi becomes decorative. Neither failure is really about knowledge. It's about delivery.
This is the exposition tightrope every hard sci-fi writer walks: too much lecture and you kill momentum, too little and the plausibility that makes the genre worth reading collapses. A reader who picked up your novel because the back cover promised orbital mechanics and reactor failures wants to feel the mechanism, not just be told it's broken. But they also don't want a physics professor cosplaying as a chief engineer.
AI models handle this tension in wildly different ways, and if you're trying to write science fiction with AI, knowing which model defaults to which failure mode saves you hours of revision. So I ran the same scene through three of them, back to back, and paid attention to exactly where each one cracked.
The Test: One Drive Malfunction, Three Models
I used an identical prompt across Claude, GPT, and Gemini, no follow-up coaching, no examples of "good" exposition fed in advance. Just the raw scene request, because that's how most writers actually work on a first pass. The setup: a ship's Alcubierre-style drive throws a containment fault mid-transit, the engineer has ninety seconds to decide whether to scram the field or ride it out, and there's a junior crew member in the room who needs to understand enough to help without turning into a walking Wikipedia article.
Write a 400-word scene aboard the freighter Kessler-Ohara. Chief Engineer Vashti Rho is diagnosing a warp containment fault — the negative energy density field is destabilizing faster than the safety margins account for. She has a junior engineer, Priya, in the room who needs to understand the danger enough to act, not just watch. Vashti needs to decide in the next ninety seconds whether to scram the field or attempt a manual re-null. Ground this in real physics extrapolation (Alcubierre metric, exotic matter, field decay), but the scene must read as a character under pressure, not a lecture. No info-dump paragraphs. All technical content must arrive through action, dialogue, or Vashti's clipped internal shorthand.
That prompt is deliberately loaded with constraints, because a vague prompt just invites each model to default to its worst habit. Even loaded, though, the differences were stark.
Claude: Character-Filtered Reasoning That Sometimes Over-Trusts the Reader's Patience
Claude's version routed almost everything through Vashti's point of view. Instead of describing the field decay abstractly, it gave her a specific tactile relationship to the ship: she reads the containment gauge as a heartbeat, she has a personal rule about scramming ("if the null drifts more than four percent, we're guessing, and I don't guess with exotic matter"), and Priya's confusion becomes the vehicle for clarifying stakes without Vashti ever "explaining" anything outright.
This is Claude's real strength in technical scenes: it tends to reason as the character rather than reasoning about the science and assigning the character a mouthpiece role. The physics arrives filtered through professional instinct, which reads as competence rather than lecture. When Vashti thinks in fragments — "four percent. maybe five. field's not linear anymore" — that's exposition doing double duty as characterization.
Where Claude stumbled: it occasionally couldn't resist finishing the thought. After a sharp, clipped internal line, it would sometimes add a clarifying sentence that spelled out what the fragment already implied. It's the AI equivalent of a good joke followed by someone explaining the joke. The instinct is generous — it wants the reader to definitely understand the stakes — but it undercuts the tension that the clipped voice just built.
If you're working with Claude on technical scenes, the fix isn't a different prompt so much as a follow-up pass. Ask it explicitly to find and cut the sentence that restates what the previous line already implied. This is worth doing as a standalone edit step, not baked into the first draft request, because Claude needs the room to overwrite before it can identify its own redundancy.
Reread the scene you just wrote. Find every instance where a clipped, fragmentary line of internal thought is immediately followed by a fuller sentence that re-explains the same information in plainer language. List those instances, then rewrite the scene cutting the redundant clarifying sentence in each case — trust the fragment to carry the meaning on its own.
This two-step approach — generate, then hunt for self-explaining redundancy — consistently tightens Claude's technical scenes without losing the internal-voice texture that makes them work in the first place. It's a pattern worth building into your standard workflow if you're leaning on Claude for any genre where a character's expertise needs to show rather than tell; the same trick works for legal thrillers, medical dramas, or anywhere professional jargon needs a human filter.
GPT: Dialogue That Snaps, Science That Softens
GPT took the opposite structural approach. Almost none of the technical content lived in internal monologue — it came out through back-and-forth dialogue between Vashti and Priya, and the dialogue had real rhythm. Short exchanges, interruptions, Priya asking the exact question a reader would ask ("Why not just kill the field now?"), Vashti answering in a way that sounded like a person talking, not a textbook.
This is GPT's signature move in high-tension scenes: it writes propulsive, naturalistic dialogue almost by default, and dialogue is genuinely one of the best exposition delivery vehicles because it disguises information as conflict. Two people disagreeing about what to do next can convey enormous technical stakes without either of them "explaining" anything — they're just arguing, and the reader absorbs the physics as a side effect.
But here's where GPT's version got shaky: to keep the dialogue snappy, it started rounding the science down. "The field's collapsing, we've got maybe a minute" is dramatic and clear, but it's also the kind of line that could describe literally any ship malfunction in any franchise. The specific mechanism — negative energy density thresholds, the Alcubierre-specific reason a scram is dangerous rather than simply cautious — got smoothed into generic urgency. It wasn't wrong, exactly. It was just no longer hard sci-fi; it had drifted toward space opera.
The fix here is a prompt that forces specificity back into the dialogue without killing its rhythm. Rather than asking for "more science," which tends to produce a paragraph break in the middle of the conversation, ask GPT to make the argument itself more specific — get the characters disagreeing about a concrete number or mechanism rather than a vague threat level.
Revise the dialogue in this scene so the disagreement between Vashti and Priya is about something mechanically specific — not "we're running out of time" but a concrete disagreement over method: does the manual re-null risk over-correcting the exotic matter density past a safe threshold, or is the scram itself what risks a catastrophic decompression event? Keep every line under 15 words. The characters should sound like they're arguing about a real decision with a real wrong answer, not describing danger in the abstract.
That "keep every line under 15 words" constraint matters more than it looks — it's what stops GPT from resolving the tension between specificity and pace by adding an explanatory paragraph. It has to fit the mechanism into the argument itself. If you write across multiple genres, this same tactic — forcing dialogue to disagree about specifics rather than gesture at stakes — is one of the more transferable tricks in best AI models for writing comparisons; it shows up as a fix for thin conflict in thriller with AI work too.
Gemini: The Info-Dump Wearing a Dialogue Costume
Gemini's version looked like dialogue on the page — quotation marks, back-and-forth structure, character names attached to each line — but functionally, it was an info-dump. Priya would ask a short question, and Vashti's "answer" would run six or seven sentences of continuous technical explanation before Priya spoke again. It's a subtler failure than a straight lecture paragraph because it's formatted like conversation, but the imbalance gives it away instantly: no real person explains a containment field collapse in one uninterrupted breath while the room is actively depressurizing.
This is a pattern worth watching for specifically with Gemini — it tends to reach for structural completeness. It wants to make sure the reader has all the information, and dialogue becomes a container for a fully organized explanation rather than a messy, interruptible human exchange. The science itself was accurate, often more precise than GPT's version, but the delivery mechanism betrayed it.
The prompt fix that works best here isn't asking for "less exposition" — that's too vague and tends to make Gemini cut technical content rather than restructure its delivery. Instead, cap the turn length explicitly and force interruption into the mechanics of the conversation itself.
Rewrite this scene so no character speaks more than two sentences before being interrupted, cut off, or responded to. If Vashti needs to convey a multi-part technical explanation, split it across at least three separate exchanges, with Priya reacting, pushing back, or acting in between each part. The ship should feel like it's interrupting them too — an alarm, a shudder, a readout changing — at least once during the explanation.
That last sentence — forcing the environment itself to interrupt — is the detail that does the most work. It breaks up Gemini's tendency to let a character hold the floor long enough to deliver a complete lesson, and it adds the kind of sensory grounding that makes a malfunction feel physically real rather than narratively convenient. If you're building out a broader fantasy worldbuilding or hard sci-fi system and using Gemini as your primary drafting model, this interruption technique is worth keeping as a standing instruction in your prompt library, not a one-off fix.
A Scoring Rubric You Can Actually Use
Comparing outputs by gut feeling gets you inconsistent revision decisions across a manuscript. What worked for chapter three might get waved through in chapter twelve because you were tired, and three months later your technical scenes read like they were written by three different people — which, if you're switching models, they were. A simple rubric fixes this. Score each technical scene on three axes, 1 to 5:
- Accuracy — Does the science extrapolate logically from real physics, or does it hand-wave past the mechanism the plot depends on? A scene can be vague and still accurate if it's vague about the right things (implementation details) rather than the wrong ones (cause and effect).
- Integration — Is the technical content doing narrative work — building tension, revealing character, forcing a decision — or is it just sitting in the scene because the plot requires the reader to know it? Exposition that could be deleted without changing anything except the reader's comprehension is integration failing.
- Voice — Does the explanation sound like it's coming from this specific character, with their specific relationship to the material, or could you swap in any competent professional and the lines would read the same? This is the fastest tell for AI-generated technical dialogue, because generic competence is the default output of every model when unprompted.
Run this rubric during your revision pass rather than while drafting — trying to score prose as you generate it just slows you down and makes you second-guess sentences that would resolve themselves in editing anyway. If you're following The Five-Pass Revision Order for AI-Assisted Novels, technical exposition scoring fits naturally into the pass focused on scene-level function, right alongside checking whether every scene earns its place.
It's also worth keeping a running list of which model handled which kind of technical content best, because the pattern holds across a manuscript. In my experience: Claude for scenes built on a single character's expertise and internal reasoning, GPT for scenes that live or die on dialogue tension between two people who disagree, Gemini for scenes with genuinely complex multi-part systems where accuracy matters more than voice — provided you build in the interruption constraint every time.
The model isn't the variable that matters most. The constraint you put around it is. Every one of these models can write good hard sci-fi exposition; none of them will do it by default.
Building This Into Your Actual Workflow
If you're drafting a full novel rather than a single test scene, it helps to lock these patterns into your story bible rather than re-explaining them in every prompt. Add a section specifically on "technical exposition rules" — cap dialogue turn length, ban explanatory internal monologue that restates the previous line, require environmental interruption during any multi-part explanation — and reference it explicitly when you prompt for a new technical scene. This is the same principle that keeps character consistency intact across a manuscript: constraints that live in a reference document survive longer than constraints buried in a single prompt you'll forget you wrote by chapter fifteen.
It's also worth running a dedicated pass with the Manuscript Cleanup Report once your draft is assembled, specifically flagging paragraphs that run long without dialogue or action breaks — that's often where lecture-mode exposition hides even after you've caught the obvious offenders. And if you're outlining a hard sci-fi novel from scratch, building your major technical set-pieces into your AI book outline early — noting which scenes need to carry hard science and which model you'll draft them with — saves you from discovering the mismatch three chapters deep.
One more thing worth doing before you commit to a model for your whole manuscript: read the best AI models for writing comparison alongside this one. Genre and scene type change which model's default habits help you and which ones fight you, and hard sci-fi exposition is one of the more unforgiving tests because the failure modes are so visible to genre-savvy readers.
The real takeaway: don't ask any model to "write hard sci-fi exposition well" and hope for the best. Ask Claude to filter the science through a character's professional shorthand and then hunt down its own over-explaining. Ask GPT to make its snappy dialogue argue about something mechanically specific instead of vaguely urgent. Ask Gemini to let the ship interrupt the lecture it wants to write. Pick the constraint that matches the model's known failure mode, put it in the prompt every time, and your drive malfunction scene will read like a crisis instead of a syllabus.