Teaching AI to Explain Games: The Hardest Part of AI Game Development

AI can build a working game in hours. Getting that game to make sense to someone who's never played it before takes considerably longer.

April 16, 2026 · Vibe Arcade

There is a consistent pattern in AI-built games, and we have seen it across every title in the Vibe Arcade catalog: the AI builds the mechanics first and the explanation last. Usually not at all. The overnight pipeline produces games that are functional, polished, and genuinely playable. But "playable" assumes you already know the genre. Hand the same game to someone who has never seen the genre before, and you get confusion, frustration, and a closed browser tab.

The Problem: AI Builds for People Who Already Know

When an AI generates a game, it draws on patterns from thousands of games in its training data. It knows what an idle clicker looks like. It knows how nonogram puzzles work. That deep familiarity is what makes the output so functional on the first pass. But the same familiarity creates a blind spot. The AI cannot model the experience of encountering the genre for the first time. It does not add a how-to-play section because, from its perspective, the mechanics are self-evident.

Except they are not self-evident to a ten-year-old who has never played an idle clicker, a parent who has never seen a nonogram, or anyone whose mental model of "browser game" starts and ends with Wordle.

Pulse: "People Don't Know What to Even Do"

Pulse is an idle clicker. If you have played Cookie Clicker or Adventure Capitalist, the interface makes immediate sense: click to generate a resource, buy generators, unlock upgrades, prestige for permanent bonuses, accumulate progress while away. Experienced players recognize the loop within seconds.

Pulse launched without any how-to-play content. The AI had built a clean interface with generators, upgrade paths, a prestige system, and offline progress. Everything worked. And the first feedback from real players was: "People don't know what to even do."

The interface presented a wall of numbers, buttons, and labels that assumed familiarity with the idle clicker loop. What is a generator? Why would I buy one when I can just click? What does prestige mean? Why did my progress reset? Every one of those questions has an answer that feels obvious to genre fans. None of them are obvious to anyone else. A how-to-play section was added after launch. It should have been there from the start. The AI never thought to include it because the AI already knew how to play.

Hanjie: The How-to-Play That Was Added, Removed, and Added Again

Hanjie is a nonogram puzzle, sometimes called picross. You fill in cells on a grid based on number clues along the rows and columns. If you have played picross before, it clicks quickly. If you have not, it is genuinely hard to explain. The relationship between clue numbers and grid cells is spatial and logical in a way that resists bullet-point instructions.

The first version included a how-to-play overlay that auto-displayed on first visit. Feedback was mixed. Some people found it helpful. Others dismissed it without reading, then could not figure out the puzzle. Returning players found it intrusive.

So the overlay was removed. The content moved to a section below the game. This was cleaner but created a different problem: new players never scrolled down far enough to find it. They stared at the grid, poked at cells randomly, and left.

The overlay came back a third time, redesigned. Click-to-open instead of auto-displaying. A whole-board worked example instead of text-only rules. It did not force itself on returning players, but it was discoverable enough for new ones. Three iterations, and the solution came from watching humans fail at the puzzle, not from the AI reasoning about what a new player would need.

Why AI Has This Blind Spot

This is not strictly an AI-specific problem. Human game developers ship games without adequate onboarding all the time. The curse of knowledge is well-documented: when you have been building a game for months, you lose the ability to see it through fresh eyes.

But AI amplifies the problem. A human developer at least had a first encounter with the genre at some point. They remember, however dimly, what it was like to not understand. An AI has no such memory. It pattern-matched across thousands of examples simultaneously. There was never a moment where it did not know how an idle clicker works. The concept of "not knowing" is not part of its experience.

The result: AI-generated games consistently optimize for the player who already gets it. The interface is clean and efficient for experienced players. The controls follow genre conventions. What is missing is the bridge for the player who does not yet get it — the explanation that transforms a wall of unfamiliar elements into a coherent experience with a clear goal.

The Fix: How-to-Play as a Pipeline Requirement

The process change was straightforward. Every new game now includes how-to-play content as a required deliverable, not something added after launch when someone complains. Onboarding is a first-class concern alongside mechanics, visuals, and scoring.

The playability rubric now includes "clarity of goal" as an explicit scoring dimension. A game can have beautiful animations and tight mechanics and still score poorly if a cold player cannot figure out what to do within thirty seconds. That dimension did not exist in early rubric versions. Adding it changed what the pipeline produces.

But here is the honest part: the AI-generated how-to-play content is often the weakest part of the output. It writes instructions that are technically accurate but assume too much context. It explains what to do without explaining why. It uses genre terminology without defining it. The how-to-play sections usually need human editing — ironic, given that the game code itself often ships with minimal changes.

The Real QA: Humans Trying Games Cold

The most reliable quality signal is watching someone play the game for the first time with no guidance. Kids are especially good for this, not because they are less capable, but because they have not internalized genre conventions. An adult gamer might figure out an unclear interface through pattern recognition. A kid who has mostly played Roblox and Minecraft will look at an idle clicker and genuinely not know what the numbers mean.

This is not a formal usability lab. It is closer to handing a phone to a family member and watching what happens. Where do their eyes go? What do they tap? When do they stop trying? That feedback loop — human, informal, impossible to automate — is what actually closes the gap between "the AI understands the game" and "a first-time player understands the game." The how-to-play content gets revised not by the AI but by a human who watched someone struggle.

What Is Still Unsolved

The pipeline requirement has improved the baseline. Games now ship with onboarding content instead of without it. But the quality still varies, and the problem is deeper than "add instructions."

Some mechanics are easy to explain in words: match three gems, type the word before time runs out. Others are fundamentally spatial or procedural — things you understand by doing, not reading. Nonograms fall into this category. For those games, static how-to-play is necessary but not sufficient. Interactive tutorials or guided first levels might be better, but generating those automatically is a harder problem than generating the game itself.

There is also the question of when onboarding becomes patronizing. Returning players do not want explanations they have already absorbed. The click-to-open pattern is a compromise. The ideal is a game that teaches through its own design, where the first interaction makes mechanics clear through play rather than reading. That is genuinely hard game design. Human studios spend months on it. We are not there yet with AI.

For now, the process is: build the game, write the how-to-play, hand it to someone who has never seen it, watch them, revise. It is not elegant. It is not fully automated. But it produces games that more people can actually enjoy rather than abandon in the first ten seconds. The AI builds the game. The humans teach it how to explain itself.