The AI is only as good as the prompt. Generic prompts make generic worlds. Specific prompts with concrete sensory detail and a clear conflict produce worlds that feel like they were already there, waiting.The AI is only as good as the prompt. Generic prompts make generic worlds. Specific prompts with concrete sensory detail and a clear conflict produce worlds that feel like they were already there, waiting.

#The three-sentence pattern

Most good prompts have three parts: setting, conflict, player role. Each one is a single sentence with concrete detail. not an adjective pile.

#Why this works

Look at what each example does:

• Concrete visual that fires an image immediately (bone-dry Texas, dragon the size of a cat, collapsed megatower).
• Specific conflict with a number, a body count, or a stake (four people this month, three days, the last elevator shaft).
• Personal stakes for the player (two weeks on the job, inherited from a great-aunt, one daughter to feed).

The AI uses each concrete detail as an anchor. A prompt with six concrete details produces a world with six genuine features. A prompt with zero concrete details produces six filler features.

#Prompt dials you can turn

Once you've got the three-sentence core, these dials let you shape the output further. You can add any of them as a fourth sentence or leave them out.

• Tone reference. "Tone: Cormac McCarthy meets Hideo Kojima." Two references mixed is better than one.
• Mechanical emphasis. "Emphasis on investigation and dialogue, combat is rare and lethal."
• NPC texture. "NPCs should be exhausted, superstitious, and slow to trust."
• Forbidden elements. "No orcs, no elves, no prophecies."
• Genre hybridization. "Visual novel pacing with tactical turn-based combat scenes."
• Moral weight. "Every major choice should have a cost. No clean wins."

#Anti-patterns (what not to do)

These prompts all produce dull worlds. They're the most common mistakes.

• Adjective pile → "A dark mysterious gritty noir cyberpunk dystopian fantasy world." Six adjectives, zero concrete details. The AI has nothing to anchor to.
• Wikipedia summary → "A medieval fantasy kingdom with knights, castles, dragons, and magic." You've described a genre, not a world.
• Mechanical shopping list → "A dungeon crawler with permadeath, procedural generation, loot drops, and character progression." The genre settings already handle this. tell the AI the world, not the systems.
• Vague player role → "You are a hero." Make the player a specific person with a specific problem. "You are the last wheelwright in a town that just banned the wheel" is infinitely better than "you are a hero".
• Asking the AI to decide → "Surprise me with something cool." If that's what you want, use the Surprise Me button. it has better prompts than anything that phrase will generate.

#Rapid iteration

A good prompt rarely gets written in one pass. The best workflow is to iterate fast and cheaply:

1. Write a three-sentence prompt.
2. Generate the world (50 credits).
3. Read the AI's first 3 scenes. Does the world feel right?
4. If yes → play it. If no → kill it, rewrite the prompt with one more concrete anchor, and regenerate.

#Prompt templates by goal

Pick whichever fits what you want to play:

• "I want to feel something" → dark prompts with specific stakes and one NPC who has something to lose.
• "I want to solve a mystery" → a prompt that ends with an unanswered question and names the investigator.
• "I want to escape" → a prompt rich in sensory detail, low on conflict, with a clear day-to-day rhythm.
• "I want to roleplay a specific character" → put the character in sentence three and let the setting exist to test them.
• "I want tactical combat" → pick the right genre (roguelike, TTRPG, dungeon crawler) and describe the battlefield, not the story.
• "I want a long campaign" → set session length to long, describe a world with factions that have goals that conflict with each other.

#Prompts for AI-built games (vibe-coded gameplay)

The arcade family routes through wilds.ai's codegen loop — an LLM agent writes the actual React component tree, enemy AI, bullet patterns, and HUD in response to your prompt. Different prompt levers apply. See AI-built games for the full loop explanation. the short version:

Name the genre clearly. "Arcade shoot-em-up where you dodge spirals" gives the agent a clear archetype and tells it which primitives to reach for. "Make me a cool game" gives it nothing to anchor on; the loop spends iterations exploring.

Name your enemies + bosses by archetype. "Crystal cultists swarm in clusters; a giant prism boss with three phases" maps directly to useSwarm2D + a 3-phase boss config. The agent can vibe-code "swarm of small enemies" + "phased boss" without you spelling out the primitives. "Enemies attack" doesn't give it enough to work with.

Declare biome + atmosphere upfront. volcanic-archipelago biome + storm weather → the loop knows to mount volumetric clouds with sulfur-tinted density floor. Vague atmospherics → loop picks defaults that may not match your vision. The 9 biome ids: temperate-forest, arctic-tundra, desert-dunes, tropical-jungle, volcanic-archipelago, underwater-coral, alien-crystal, urban-rooftop, subterranean-cavern.

Pick a difficulty curve in the prompt. "Gentle wave ramp for casual players" vs "punishing scaling, all 5 archetypes by wave 5" maps directly to useEnemySpawner2D({ curve: 'gentle' }) vs { curve: 'punishing' }. The three presets cover most cases. you can also describe a custom curve in plain English and let the agent translate.

Describe what should happen, not the implementation. "Hits explode in colored sparks" tells the agent to use particle bursts. "Use ParticleBurst2D with count=30" is over-specified — the agent already knows the library, and the judge tracks score the effect, not the technique.

Reference compound behavior explicitly. "Enemies chase the player but retreat below 30% HP" tells the agent to build a Selector BT with BTWaitForHpThreshold + BTFlee. "Smart enemies" leaves it ambiguous.

Avoid cloning real games. "Like Geometry Wars but with biome-tinted particles" is fine — it gives a genre anchor. "Make me Geometry Wars exactly" is not — the judge tracks penalize derivative work, and copyrighted aesthetics will be auto-flagged.

A working prompt template for arcade vibe-coding:

Genre: top-down shoot-em-up. Biome: volcanic-archipelago at dusk with storm clouds. Player: a glass spectre that dashes between obsidian pillars. Enemies: lava imps that swarm in clusters of 5, ash phoenixes that orbit slow then dive-bomb, basalt sentries that patrol fixed lanes and stop to fire aimed bullets. Boss: a molten kraken with three phases — phase 1 is radial bullets, phase 2 adds patrolling tentacles, phase 3 summons imp waves. Difficulty: medium ramp, 4 waves before boss. Aesthetic: sulfur particles, screen-shake on heavy hits, vignette + bloom for that 90s arcade glow.

Every concrete detail in that prompt maps to a specific primitive the codegen loop will reach for.

#The fastest way to get better

Write three prompts this week. Generate all three. Play each one for 15 minutes. Then write a fourth prompt that takes the best detail from each of the first three.

You'll feel yourself getting better. The AI isn't changing. you're learning which levers matter.

#References

1. Getting Started: The three-minute quickstart.
2. Your First Hour: Guided walkthrough using these prompt patterns in context.
3. Creator Studio: Edit and refine the world the AI generates from your prompt.
4. Session Tactics: How prompt quality affects session depth.
5. Browse and Play: See what prompts other creators used.