Building a Game Demo in One Day with Claude Code: A Full Retrospective

One Sunday afternoon, a developer imported a game template. The time was 1:46 PM. When he finally looked up again, it was 3:00 AM, and the last commit message read "trial (demo)". In between: 86 commits.

The result was a game called Reptile Catcher: you head out to stun and catch lizards, sell them, and breed them for more exotic patterns. It runs on the Godot engine, with code written by Claude Code. The developer's job was to set the direction, do playtesting, and select art assets.

The most valuable part of this retrospective isn't "AI can write code" — that's old news — but rather what it reveals about how the human's role shifts when collaborating with AI.

Core Loop Running in Five Minutes

Five minutes after importing the template, the "catch-sell-breed" core loop was up and running. The developer admitted this was expected — today's AI handles this kind of work easily.

What really mattered was what happened in the next minute. The very next commit was a balance simulator that lets a bot play the game for you.

Incremental games are fundamentally a numbers game. Change one upgrade price and a 30-minute flow can stretch to 3 hours. The developer couldn't replay the whole thing from scratch every time, so whenever Claude adjusted a value, he'd run the bot and make decisions based on data like "reached the ending in 4 hours 19 minutes (bot)." They later split the bot into five personalities: one prioritizes buying attack upgrades, one only buys the cheapest option, one only breeds, and so on.

The bot also caught a bug in the process. While measuring the demo's playtime, one bot would never finish — it was stuck trying to purchase an upgrade that was locked in the demo. The bug was on the bot's side, but it exposed the fact that locking rules were scattered all over the codebase.

Building the Simulator First Was the Best Decision of the Entire Project

The author offers a clear piece of advice: if you're using AI to build a numbers-heavy game and don't have a simulator yet, consider having the AI build a bot right after the core loop is done. For him, it was the single best move in the entire project. The lesson generalizes well — AI is great at piling up code fast, but humans need quantifiable feedback to make decisions. The simulator is the bridge that converts "subjective feel" into "objective data."

What is an Incremental Game? An incremental game is a genre centered on numerical accumulation. Players repeatedly perform simple actions to collect resources, then spend those resources on more efficient upgrades, creating a positive feedback loop. These games are extremely sensitive to balance: upgrade curves that are too steep make players feel stuck; curves that are too flat eliminate the satisfaction of progression. Traditionally, designers had to manually replay the game repeatedly to sense the pacing — extremely time-consuming. A bot simulator addresses this by scripting the decision logic of different player styles, running hours of gameplay in seconds, and converting the subjective experience of "feel" into quantifiable metrics like "completion time" and "resource curves." Splitting into five bot "personalities" further simulates the behavioral diversity of a real player base, surfacing balance issues that a single test path might miss.

2:20 PM — Starting Over

What the first 30 minutes produced was a screen full of tabs and buttons. By 2:20 PM, the developer had completely changed direction: switched to pixel art, turned "catching" into a hands-on chase level where you physically move your character, and redesigned upgrades into a skill tree that branches outward like a mind map.

Normally, throwing everything away at this point would feel terrifying. But the cost to rebuild was so low that he barely hesitated. This is the biggest difference he felt when working with AI — the marginal cost of trial and error has been pushed to near zero, making directional pivots no longer expensive decisions.

AI Generates Sprites, Humans Do the Selecting

Pixel sprites came from PixelLab, integrated into Claude via the MCP protocol. Claude generated several candidates; the developer picked one and left notes like "go with #3, but shorten the neck a bit." Claude recorded the seed of each selected sprite in a document.

Interestingly, some species were swapped out for specific reasons:

• Geckos shrank to an indistinguishable shape at 32 pixels in side view, so they were replaced with skinks;
• The gliding lizard's wing membrane was invisible at 32 pixels, so it was replaced with a frilled lizard;
• The bearded dragon's walking animation always looked like a small dog no matter the adjustments, so it was replaced with a horned lizard.

The developer didn't obsess over generating images until he got exactly what he had in mind. Instead, he adjusted the game to fit what the tool was good at drawing. For a small game, that's much faster. This tradeoff thinking is worth reflecting on: rather than forcing AI to produce your ideal mental image, it can be smarter to redesign the requirement around the tool's strengths.

What is MCP? MCP (Model Context Protocol) is an open protocol introduced by Anthropic that lets AI models connect to external tools and data sources in a standardized way. By connecting PixelLab (a pixel art generation service) to Claude via MCP, the developer could call the image generation API directly during a conversation, without manually copying and pasting between different tools. A "seed" is the random number initial value in an image generation model — the same seed combined with the same prompt reproduces the same image. Recording seeds in a document means you can later make precise tweaks based on a previously selected image, rather than generating from scratch again. This workflow turns AI image generation from "randomly hoping for the best" into a traceable, iterable asset management process.

My Job Was to Delete What AI Built

Around 7:00 PM, the developer deleted an entire suite of features Claude had built that afternoon: auto-catch, auto-sell, and offline progress. Ten minutes later, the prestige system and prestige tree were cut too.

He had established the game's tone early: "no failure, no prestige resets, a single linear path, 4 to 6 hours to complete." Idle mechanics and prestige systems had no place in that framework. AI builds fast and features pile up quickly, but deciding what to cut is still entirely up to the human. Looking back, that was essentially all he did on day one.

This reveals an interesting reversal: in the AI era, a developer's core value may no longer be about "building" — it may be about "deleting" — maintaining directional restraint and consistency amid an explosion of generated features.

What is a Prestige System? Prestige is a common design pattern in incremental games: once the game reaches a late stage, players can choose to "reset" all progress in exchange for permanent multipliers or newly unlocked content, starting a new, more efficient cycle. This mechanic significantly extends game lifespan and is a core retention mechanism in long-running incremental games like Cookie Clicker and A Dark Room. But prestige requires players to be willing to invest enough time and accept the psychological cost of wiped progress. For a game designed as a "4 to 6 hour linear narrative," introducing prestige would fundamentally break the "single straight-line flow" design intent — leaving players confused about whether they should reset, while diluting the completeness of a single playthrough. That's exactly the logic behind the developer's decision to cut it.

One Day Is Just the Demo — Shipping Still Took a Week

By 3:00 AM, the demo's scope had converged — but the actual release date was October 11th. In the time between, Claude added four bosses, drew scene backgrounds, fixed the demo's locking bug, and prepared screenshots, cover art, and a web build for the itch.io page.

So the claim "built in one day" only holds at the demo level. Building took one day; polishing and shipping took a week. The developer also admitted the next game will probably follow the same rhythm.

This honest ending captures the real picture of AI-assisted development: the acceleration from 0 to 1 is genuine, but the "last mile" from 1 to shipped still requires considerable human time. Behind the dazzling "one-day product" is the reality of a week of polish.

The demo is live on itch.io — go catch some lizards 🦎 👉 https://krenin.itch.io/reptile-catcher