Build the world,
not one prompt.

Sam constrained the live build to a 2D/2.5D isometric top-down game because polished 3D would spend too much of the hour on models, canvases, and external assets. The chosen brief became low-poly isometric tiles, a procedurally generated world, camera pan and zoom, settlement growth, villagers, and lighter logistics so it could become playable quickly. A reference can guide structure, but copying another game’s protected assets is not acceptable.

The cleaned first prompt was: I want to build a 2D game. At the session date, explicitly naming a 2D, 2.5D, or 3D game routed most users into the newer backend. Expanding the builder’s running read status and seeing a game/ prefix was the diagnostic clue. The older path could still build games, but Sam said it required more hand-holding.

Sam connected a Super Agent to the blank app, either by same-workspace app name or an API key, and asked it to turn plain spoken requirements into structured builder prompts. His role instruction was effectively: You will be my builder as a Super Agent; I will speak to you in plain-text English and you will formulate the prompts. ChatGPT, Claude, or Gemini could clean the thinking too; the Base44 agent removed the copy-and-paste step.

A flat AI image or a request to “make it look better” is not a production art pipeline. For 2D and 2.5D work, use sprite sheets containing the frames for an object or character state. Generate or source a reference, convert it into compatible sprite or model assets, upload and tag it, and keep it replaceable. Programmatic motion is often better for water and weather, while sprite sheets suit characters, creatures, trees, and authored idle states.

Sam designed a backend for reference uploads, environment and character assets, sprite sheets, model and animation generation, tags, tree sway, NPCs, weather, placeholders, and LLM context. The owner becomes the orchestrator and can replace content without repeatedly returning to the builder. File uploads were the intended source of truth, with AI helping interpret and prepare them.

Creating an Asset Studio does not tell the game to consume it. Sam separately prompted the front end to use approved generated assets and replace the placeholder trees with the new tagged oak-tree asset and animation. He later used the same pattern for weather: Please ensure the front end is hooked up to our backend weather system.

The generated tree initially had an unwanted circular base and a non-transparent background. Sam corrected the pipeline with a transparent-background constraint and recommended screenshots in debugging prompts. A production studio should keep multiple variations—he suggested up to roughly 10—then randomize variant and size so repeated scenery does not look mechanically identical.

Treat the first output as a proof of concept. Sam’s preferred order was levels, NPCs, automations, save state, movement, menus, and other fundamentals before production visuals. Lighting, shadows, and final art come after the systems work, but the game should not reach users until it looks intentional.

Sam recommended WebSockets for live multiplayer rather than timestamp polling. First create a public landing page, required login, a unique username or character profile, and user-specific game state. Then request a WebSocket multiplayer experience so each player keeps an independent session while seeing other players live.

Clouds, rain, snow, fog, and daylight can be programmatic or asset-driven, but the front end must be connected to the backend weather system. Sam’s illustrative prompt used a six-hour weather system, a 24-hour day/night clock, and a 20:1 ratio where one real second represents 20 in-game seconds. Those numbers were examples, not platform limits.

Sam named Meshy as an example 3D-asset provider. The proposed workflow was to obtain an API key, store it as a Base44 secret, and tell the builder which provider and library to import. Base44 can assemble models, maps, and animation packs, but a text-only LLM prompt should not be expected to create cinematic GTA or Cyberpunk-quality art.

Sam’s ideal template would include a story and narrative section that every later generation reads before making decisions. That persistent world context would stop tone-breaking content—for example, cute assets appearing in a cyberpunk setting—and would sit beside graphic-style constraints, animation rules, sprite sheets, and the asset inventory.

There is no universal credit total. The one-prompt 2D Space Invaders demo cost 8.4 credits; Sam projected 40–60 for the limited workshop build and 1,000–1,500 for his desired procedural, NPC, WebSocket co-op version. He estimated three to five hours for a one-map prototype and hundreds of hours for a polished multilevel multiplayer release. These were personal examples, not pricing guarantees.