Tower Before Dusk: How I Built a Puzzle Game Designed for Both Humans and AI to Play

A twilight tower puzzle game is a type of logic-based indie game where players stack, sequence, or navigate structures under a time-or-light constraint — designed here to be equally solvable and enjoyable by human minds and AI systems.

Why I Decided to Build a Puzzle Game for Humans and AI

When most indie developers talk about their audience, they mean players — human beings who sit down, pick up a controller or keyboard, and lose themselves in a world. I had a different idea. What if the game itself was a bridge? What if every mechanic, every puzzle rule, every visual cue was legible not just to a person, but to a language model or a reinforcement learning agent trying to reason through the same challenge?

That question became Tower Before Dusk — a minimalist puzzle game built over four months as a solo developer, using Godot 4 and a deliberate design philosophy I started calling dual-audience design.

What Is Tower Before Dusk, Exactly?

At its core, Tower Before Dusk is a turn-based stacking puzzle. You place geometric blocks on a tower before an in-game sun sets — roughly 90 seconds of real time. Each block has a weight value, a color-coded stability rating, and a placement rule. Stack incorrectly and the tower collapses. Stack perfectly and you unlock the next level’s blueprint.

The twist: every game state can be represented as a simple JSON object. Block positions, weight distributions, remaining time — all of it is machine-readable by design. I built a lightweight API layer into the game so that an AI agent (or any developer with API access) can query the state, reason about it, and submit moves just like a human would through the UI.

Within two weeks of launching a public demo, three separate developers had connected GPT-4 and Claude-based agents to the API and posted their results on a game dev forum. One AI agent cleared Level 12 — something only about 8% of human players had done at that point.

The Design Philosophy: Rules That AI Can Read, Humans Can Feel

1. Explicit, Enumerable Rules

Most puzzle games rely on intuition — the “feel” of a physics engine, the visual rhythm of a platformer. Those are nearly impossible for current AI to internalize without massive training data. So I made every rule in Tower Before Dusk explicit and enumerable. There are exactly 7 block types, 3 stability states, and 4 placement zones. Any AI with basic reasoning can parse a move’s validity in milliseconds.

2. Emergent Complexity From Simple Axioms

Simplicity in rules does not mean simplicity in challenge. By Level 20, the interplay between weight ratios and time pressure creates genuinely hard optimization problems — the kind that stump both humans and greedy-algorithm AI agents. According to a small internal test with 40 players, average completion time for Level 18 was 11 minutes for humans versus 3.2 seconds for a well-prompted GPT-4 agent — but the AI failed 60% more often on Level 20 due to long-horizon planning gaps.

3. A Shared Visual Language

The color system in the game (green = stable, amber = warning, red = critical) mirrors traffic-light conventions that appear frequently in AI training data. This was intentional. When a vision-capable AI model looks at a screenshot, it can immediately infer the game state without custom training. Human players, meanwhile, find the system intuitive from the first minute of play.

Monetization: Turning Dual-Audience Design Into Digital Income

Here is where things got interesting from a business perspective. I launched the game on itch.io with a pay-what-you-want model and a $4 suggested price. In the first 30 days, the game earned $1,840 — respectable for a solo release. But the bigger opportunity came from the API.

I introduced a Developer API Tier at $9/month, giving AI researchers and hobbyists programmatic access to game states, a leaderboard for AI agents, and weekly “AI Challenge” puzzle packs. Within 60 days, 112 developers had subscribed. That is a recurring $1,008/month from a side project that took four months to build.

The lesson: when you design for two audiences, you open two monetization channels. Human players pay for the experience. AI developers pay for the playground.

Looking for more tips on ai & digital income? Visit SAVYX

What I Learned About Building AI-Compatible Products

Keep State Serializable

If your product’s core state — whether it is a game, a workflow tool, or a content system — can be fully described in a structured format like JSON or YAML, you have built something an AI can work with. That is a massive competitive advantage in 2025.

Document Your Rules Like an API

I wrote Tower Before Dusk’s rulebook as if it were technical documentation. Short sentences. Numbered constraints. No ambiguity. This helped both human players learn faster and made it trivial for developers to write system prompts that let AI agents understand the game in seconds.

Build a Community Around the Intersection

The most engaged users were not hardcore gamers or hardcore AI researchers — they were people curious about both worlds. A Discord server for Tower Before Dusk grew to 340 members in 90 days, driven almost entirely by people sharing clips of AI agents playing alongside humans.

The Bigger Picture: AI-Compatible Creative Work Is the New Niche

Tower Before Dusk taught me that the most valuable creative products of the next decade will not just be consumed by humans. They will be co-experienced, analyzed, and extended by AI systems. Developers, writers, and designers who build with that dual audience in mind will unlock income streams, communities, and distribution channels that purely human-facing products simply cannot reach.

The dusk is not the end of the day. In this case, it was just the beginning.

Frequently Asked Questions

What is Tower Before Dusk and who is it for?

Tower Before Dusk is a minimalist turn-based stacking puzzle game designed to be played and understood by both human players and AI agents. It targets indie game enthusiasts, AI researchers, and developers interested in human-AI co-play experiences.

How can an AI actually play a puzzle game like this?

The game exposes a lightweight API that serializes the full game state into a structured JSON format. AI agents — including LLM-based systems and reinforcement learning bots — can query this state, reason about valid moves, and submit actions programmatically, just as a human would through the game’s UI.

How much money can you make from an AI-compatible indie game?

Results vary, but the Tower Before Dusk case study generated $1,840 in its first 30 days from a pay-what-you-want model, plus over $1,000 per month in recurring revenue from a Developer API subscription tier — showing that dual-audience design can create multiple income streams.

What tools were used to build Tower Before Dusk?

The game was built using Godot 4, a free and open-source game engine popular among indie developers. The API layer was a lightweight custom integration built to serialize game state and accept move submissions from external clients, including AI agents.

Can I apply dual-audience design to products other than games?

Absolutely. The core principle — making your product’s state serializable, your rules explicit, and your interfaces machine-readable — applies to SaaS tools, content platforms, educational apps, and workflow software. Any product that an AI can interact with programmatically benefits from this approach.