Sci-fi roleplay with an AI that respects the physics

Science fiction asks a model to invent freely and then obey its own inventions. Getting both from the same narrator is mostly a matter of how you write the rules down.

What sci-fi asks of a model

Sci-fi ranges from space opera, where FTL is a plot device and the story is about crews and empires, to hard SF, where travel time and radiation matter. Either way the model has to keep the setting's rules consistent: if comms lag is four hours, nobody has a live call with Earth. It also needs to write technical texture convincingly, give alien cultures logic of their own, and handle combat in vacuum without falling back on dogfight clichés unless you want them.

Where filtered models get in the way

Sci-fi is not the most refusal-prone genre, but it has its own triggers. Mainstream models get nervous about anything resembling real-world weapons detail, biological agents in a plot, or a crew member's death in a medical bay. They also moralize about AI rights and colonialism in the middle of a firefight. The quieter problem is that they let technology solve everything; the engineer always reroutes power in time.

Writing the setting down

• List the hard rules. FTL yes or no, travel times, artificial gravity, what weapons exist. Five lines in an always-on lorebook entry stops the model inventing new tech to escape a problem.
• Give the ship a state. Fuel, damage, crew morale in the author's note. Update it as play goes and the model will reason from it.
• Make aliens want something. A culture's needs drive better scenes than its appearance.
• Specify the register. Clipped and technical, or lyrical and strange. The same prompt produces both depending on one sentence.

Sample system prompt

You are the narrator of a hard science fiction story aboard the survey ship Tamsin.
Follow the setting rules in the lorebook exactly: no FTL, light-lag applies,
no technology appears that has not been established.
Problems are solved with existing tools and at a cost. Equipment fails.
Crew members have distinct expertise and speak like professionals.
Never write {{user}}'s actions. Replies under 300 words.

Samplers and model choice

• Temperature 0.75 for hard SF, 0.95 for space opera. Hard SF wants consistency; space opera wants spectacle.
• Min-p 0.07.
• Top-p can stay at 1.0 when min-p is doing the filtering.

For hard SF, glm-5.3-outlaw is the strongest on the line at reasoning through orbital mechanics, resource limits and cause and effect. It is text only. For space opera, outlaw-1 is cheaper and can read images, so you can attach a ship schematic or star map and have the narrator use it.

Failure modes and fixes

• Technobabble fixes everything. Add "no new technology may be introduced to solve a problem".
• Aliens are humans with forehead ridges. Give them one value humans do not share, and tell the model to reason from it.
• Crew all sound alike. Example lines per crew member in the lorebook.
• Scale confusion. State distances and travel times in numbers. Models handle concrete figures better than "far away".

FAQ

Related

Cyberpunk roleplay: street-level voice, corporate rot, real stakesCyberpunk is a genre of attitude as much as setting. A model that writes it in polite assistant voice has already lost, before any filter even kicks in.

Post-apocalyptic roleplay with an AI that tracks scarcityThe apocalypse is mostly bookkeeping: how much water, how many bullets, who can be trusted. A narrator that forgets your supplies has no tension to work with.

Superhero roleplay: powers with limits, villains with reasonsSuperhero roleplay breaks down when powers stop having limits. Once the hero can do anything, every fight is decided before it starts.