Every prompt you type tells an AI agent how to spend energy — reading files, writing code, generating art, running tests. Small prompts spend a spark. Big, structured prompts can build a whole game in one shot. This handbook walks you from your very first prompt to pasting an entire pitch deck.
The AI agent isn't magic. It reads, thinks, writes, and tests — and every step costs a little energy. Better prompts = less wasted energy = faster, better games.
Agent opens the files it thinks it needs. Vague prompts make it open too many.
It plans the steps. Structured prompts = shorter plans.
It edits files or generates art. Every image and every edit costs energy.
It runs the game in its head (and sometimes for real) to check it works.
These are one-line prompts. They change one thing. They cost almost no energy. Start here every time you're not sure.
Make the background blue.
The agent reads one file, swaps one color token, saves. Almost no thinking, almost no energy.
Add a button that says "Play".
Agent inserts a button element and basic styling. No state, no logic.
Change the title to "Dragon Quest 9000".
Text-only edit. Cheap, fast, safe.
Make the player jump higher.
Agent has to find the jump constant inside the game logic, change it, and confirm. Tiny code reasoning.
Center everything on the screen.
Layout reasoning — agent picks flexbox/grid. Watch for vague pronouns ("everything").
Now you're asking the AI to generate assets, add behaviors, or wire small systems. More energy, more impact.
Generate a pixel-art knight holding a glowing sword, side view, transparent background.
Agent calls the image model. New PNG is created, saved, and imported. Energy spent: image GPU + a couple of file edits.
Add an enemy that walks back and forth on a platform and damages the player on touch.
Agent writes new entity logic, collision check, damage handler, and probably edits 2–3 files. Multiple steps of reasoning.
Add a score counter in the top-right that goes up by 10 every time I collect a coin.
Agent must add: a coin entity, a collision rule, a score state variable, and a UI element. Four small jobs in one prompt.
Make the music change when the boss appears.
Agent needs an audio system, an event trigger, and a state machine for "boss fight". This is where guessing starts — give it the file/scene name to save energy.
Bullet points. Numbered rules. References to assets you've starred. These prompts feel longer but they actually save energy because the AI stops guessing.
Build a 3-level platformer where the hero is a robot cat collecting batteries. Each level should: - Use the "platformer" template - Have 8–12 batteries to collect - End with a door that only opens once all batteries are collected - Get harder: level 2 adds moving platforms, level 3 adds a boss Use the asset style from my starred concept art.
Now the agent is doing real planning. It will likely scaffold scenes, generate art, wire collisions, test, and report back. Energy ≈ a campfire of reads, writes, and re-runs.
Refactor the inventory system so items are stored in an array and rendered in a 3x3 grid. Keep all existing items working. Add a "drop" button that removes an item.
Agent reads multiple files, plans a change, edits 3–5 files, and re-tests. "Keep all existing items working" forces the AI to be careful, not creative.
Read the current GameConfig and add a new scene called "haunted_forest" that reuses the platformer template but uses my "moonlit pine trees" image for the background.
Anchored to your existing config + asset. The AI doesn't invent — it extends.
You can paste a whole PRD, a slide deck, or a real conversation. The AI reads it once, plans the entire build, and ships in one go. This is how pros work.
# Robo-Cat: Battery Quest — PRD Audience: ages 10–14 Goal: 3-level platformer, ~5 minutes to play through. Player promise: "Be a clever cat-robot who saves the city's power grid." Mechanics: - Move: left/right + jump (variable height) - Collect: batteries (worth 10 pts), gold batteries (worth 50 pts) - Lose condition: 3 hits from enemies - Win condition: collect all batteries + reach the door Level design: 1. Tutorial junkyard — flat ground, 8 batteries, 1 dog enemy 2. Rooftop city — moving platforms, 10 batteries, 3 drones 3. Power plant — boss fight: the "Old Generator" Tone: bright, optimistic, retro pixel-art with neon highlights. Out of scope (v1): multiplayer, save files, mobile controls.
Paste this whole PRD as a single prompt. The agent will read it once and produce a multi-step plan covering scenes, assets, mechanics, and UI. Energy is large up front, but TINY compared to typing 30 small prompts to get the same thing.
Slide 1 — Title: "NEON DELI" Slide 2 — The hook: "Make sandwiches before the city wakes up." Slide 3 — Player promise: cozy, fast, slightly ridiculous. Slide 4 — Core loop: take order → grab ingredients → assemble → serve. Slide 5 — Art direction: cyberpunk-meets-Studio-Ghibli, warm neon. Slide 6 — 3 levels: Quiet Tuesday, Rush Hour, The Mayor's Order. Slide 7 — Win/lose: tips earned vs. patience meter.
The agent treats slides as a structured brief. It will infer scenes from slides 6–7, art from slide 5, and the player feel from slides 2–3. Great for students who think visually first.
Me: I want to make a game about a kid who can talk to ghosts. Friend: cool, what does the kid DO with the ghosts? Me: helps them finish unfinished business so they can rest. Friend: like a quest giver? Me: yeah! each ghost has one small puzzle. solve it, ghost disappears, you get a memory. Friend: how does the player win? Me: collect all 7 memories before sunrise. Friend: art? Me: soft watercolor, mostly purples and warm yellows for the ghosts.
Messy human conversation IS a prompt. The agent will read the whole thing and extract: genre (puzzle/adventure), mechanic (7 ghost-quests), win condition (sunrise timer), art (watercolor purples). Don't be afraid to paste your real thinking.
Tape this above your laptop.