A Game Made by AI: How to Tell a Playable One From a Demo

Short answer: a demo proves the mechanic runs. A game makes the mechanic matter. That difference is decided before the first prompt, by the shape of the idea, and it can be graded in about six minutes on four axes — one repeatable verb, a number the player watches, a cost for failing, and a reason to start again. Score the idea out of eight. Six or above and the build will read as a game even when it looks rough. Three or below and no amount of art, levels, or polish will move it, because polish is not the missing ingredient.

The grade is the useful part. Most people run it backwards: they build, look at the result, feel that something is off, and start adding content. Content is the one thing that cannot fix a low-scoring idea. It only makes the demo longer.

Two wrong baselines

Judging an AI build is hard because both obvious reference points are broken.

Compare it to a finished commercial game and everything is a demo. The comparison was decided before you looked at the screen; there is no version of a two-hour build that survives it.

Compare it to the demo on the generator's own landing page and everything looks done, because those demos were hand-picked and hand-tuned. You are comparing your first draft to someone else's hundredth.

The only baseline that carries information is internal: does this build survive its own second minute, with a player who was not told what to do?

That question splits cleanly in two, and the split is the whole article.

Read the bottom row again. Three of the six lines above are settled before a single prompt is typed. That is why post-build fixing has such a poor hit rate: by the time you can see the problem, the decision that caused it is hours behind you.

The four axes, scored before you build

Each axis gets 0, 1, or 2. The scoring rules are written so that a stranger can score your idea without asking you a question, which is the only way the instrument stays honest when you are the one being graded.

Axis 1 — one verb, repeatable under rising pressure

The verb is what the player does every five seconds. Not the theme, not the setting, not the genre label. Dodge. Place. Time. Sort. Aim. Spend.

• 2 points: one verb, and the same verb holds at three intensities. Early on you dodge one slow thing; later you dodge six fast things; the input never changes.
• 1 point: the verb exists but pressure is flat. Minute five asks for exactly what minute one asked for.
• 0 points: several verbs with no shared resource. You walk, then you talk, then you fight, then you craft, and none of them feeds the others.

The test: name the verb in one word, then describe what makes it harder at minute one, minute three, and minute five. If the answer to all three is "more of them", you are at 1, not 2 — more is not pressure, it is duration.

Axis 2 — a number the player watches

• 2 points: one number that both the player and the system act on. The player optimises it; the system reacts to it. Score that raises enemy speed. Height that changes what spawns. Time left that changes the music.
• 1 point: a number exists but nothing responds to it. A score in the corner that only accumulates.
• 0 points: the goal is a state, not a number. "Escape the room." "Find the ending." "Survive the night." States end; numbers pull.

A number is what turns a scene into a loop, because a number can be improved without the content changing. This is the single highest-leverage line on the sheet, and it is also the line AI builds fail most often: generated games love a state goal, because a state goal is what a paragraph of prose describes.

Axis 3 — a price for failing

• 2 points: failure takes something measurable and the restart is under two seconds. Measurable means a number moved: a life, a run's score, a rank, a streak.
• 1 point: failure exists but costs nothing, or the restart takes more than five seconds of menus.
• 0 points: no fail state at all. The build cannot be lost, so nothing in it can be won.

The restart time is not a detail. A two-second restart makes a hard rule feel fair; a twelve-second restart makes the same rule feel like punishment, and players quit on the second one while blaming the difficulty.

Axis 4 — a reason to start a second session

• 2 points: a different run is reachable. A seed, a branch, an unlock, a score to beat, a route you did not take.
• 1 point: the same run again, with a better number, and nothing else new.
• 0 points: a one-time reveal. Once you have seen it, there is nothing left.

This axis is a gate, not a to-do list. If it scores 0, the finding is that you built a scene, and the fix is a different idea — not a hook bolted onto the ending.

The bands

The 4–5 band is where most AI builds land, and it is the band that gets mismanaged. A 4 looks close enough to finished that the instinct is to add levels. But a 4 means one axis is at zero, and a zero axis is structural: adding content to a build with no fail state produces a longer build with no fail state.

IDEA: ____________________________________________

A1  ONE VERB                                 0 / 1 / 2
    same verb, three intensities ................. 2
    verb exists, pressure flat ................... 1
    several verbs, no shared resource ............ 0

A2  A NUMBER THE PLAYER WATCHES              0 / 1 / 2
    one number, player and system both act ...... 2
    number exists, nothing responds ............. 1
    goal is a state, not a number ............... 0

A3  A PRICE FOR FAILING                      0 / 1 / 2
    measurable loss, restart under 2s ........... 2
    no loss, or restart over 5s ................. 1
    no fail state ............................... 0

A4  A REASON FOR A SECOND SESSION            0 / 1 / 2
    a different run is reachable ................ 2
    same run, better number, nothing new ........ 1
    one-time reveal ............................. 0

TOTAL ___ / 8        BAND: ______________________

Six minutes, one idea, no tools open. If you cannot finish a line, the line is the finding.

Three ideas, scored

Pitches are the only thing you have before you build, so the instrument has to work on a sentence.

"A lighthouse keeper who must keep the ships safe through a storm." A1: what is the verb? Tending the light, maybe, but the pitch describes a situation, not an action — 0. A2: no number — 0. A3: ships sink, which is a state, not a price the player pays — 0. A4: the storm ends — 0. Total 0. This is a scene with atmosphere. It is the kind of pitch that produces a beautiful, unplayable build, and the beauty is what makes people defend it.

"You dodge falling rocks; every ten dodges the rocks fall faster and a second stream opens." A1: dodge, three intensities — 2. A2: the dodge count changes the spawn rate — 2. A3: one hit ends the run, restart is instant — 2. A4: score to beat, and the second stream changes the pattern — 2. Total 8. Note how little the pitch contains: no theme, no characters, no art direction. Everything that made it score is a rule.

"Sort incoming mail into the right bin before the timer runs out; wrong bins cost three seconds." A1: sort, with the timer shortening — 2. A2: the timer — 2. A3: time is the price, and time is measurable — 2. A4: only a better score, nothing new opens — 1. Total 7. A game with one weak seam, and the seam is stated: replay is a number chase, so the build will hold for three sessions and not thirty.

The pattern across the three is consistent enough to use as a rule of thumb: pitches that describe a place score low; pitches that describe a decision under pressure score high. The word count is irrelevant. The second pitch is shorter than the first and eight points stronger.

Six idea shapes that always ship as demos

These are not genres. Genre is nearly irrelevant to the verdict — a platformer can score 8 and a shooter can score 2. What matters is the shape of the idea, and these six shapes fail predictably.

1. The content-shaped idea. The fun lives in the assets: a hundred enemies, forty levels, a rich world. The rule between the assets is thin. Grading: A1 often 1, A2 usually 0. Tell: when you describe it, you list things rather than actions.

2. The systems idea. Three resources, a crafting tree, a reputation meter, and no single number on screen that the player is actually chasing. Grading: A2 at 0 despite looking sophisticated. Tell: you need a paragraph to explain what you are trying to do at any moment.

3. The reveal idea. The pleasure is a one-time surprise — a twist, a transformation, an ending that recontextualises the build. Grading: A4 at 0 by construction. Tell: your pitch contains the word "eventually".

4. The second-player idea. The design assumes another human. Alone, the build is a waiting room. Grading: A4 at 0, A3 usually 0. Tell: any sentence containing "with a friend" as the reason it is fun.

5. The author's-head idea. The fun depends on context that never made it into the build: the joke you had in mind, the reference, the mood of the moment. Grading: A1 at 0 because the verb is ambiguous even to you. Tell: someone plays it, shrugs, and you think "they did not get it".

6. The genre-label idea. "Make me a roguelike." "Make me a battle royale." A genre is a bundle of conventions, not a verb, and a generator asked for a genre will reproduce the conventions and skip the tension. Grading: A1 at 1 at best. Tell: your prompt has a genre word in the first six words and no verb anywhere.

Every one of these can be converted. None of them can be converted by generating again with a longer prompt, because the missing axis is not a detail the model forgot — it is a decision the pitch never made.

The rewrite ladder: raise the score without adding scope

Five transformations, each one moves a single axis, none of them adds content. Apply the one that targets your zero.

Two rules for using the ladder.

Rule one: one rung per rewrite. Changing two axes at once makes the next grade unreadable, because you cannot tell which change moved the score. One rung, re-grade, then decide.

Rule two: the rewrite goes in the pitch, not the code. If you find yourself editing a build to implement a ladder rung, you have skipped the point. The ladder exists to be applied before the next generation, in the sentence you are about to type.

A converted idea usually gets shorter. "A game about a deep-sea diver exploring a wreck" becomes "you surface before air runs out" — fewer words, more game. That shrinkage is the signal that you moved a structural axis rather than decorating one.

The twelve-signal field test, after the build

The grade predicts; this confirms. Twelve observable signals, each with a pass threshold, each checkable without reading the code. Run them in order and stop counting at the first four failures — a build that fails four signals is a demo, and the rest of the list will not change the verdict.

Signals 5, 6, 7, 8 and 12 are the ones AI builds fail most, and they are all rules rather than rendering. That is the useful part of the list: the failures cluster in a place that is cheap to change and invisible in a screenshot.

The minute-two protocol

Hand the build over, say nothing, start a timer, and write down two numbers: the second at which the player stops acting deliberately, and the second at which they stop entirely. A demo has one number; a game has neither, because the session ends when the player decides to end it.

Do not explain anything during the test. Every sentence you say to help is a signal that failed. The most common false pass in this whole exercise is a build that works only because the author was narrating it.

Four signals that are not verdicts

Art quality, audio, content volume, and bug count do not appear in the twelve, and their absence is deliberate. All four affect how a game feels; none of them determine whether it is one. A build with placeholder rectangles and a priced failure is a game. A build with generated sprite art and no fail state is a demo with better lighting. Fixing the four will not move a 3 to a 6, which is why the order of work matters more than the amount of it.

What the verdict does not tell you

Four limits, stated plainly so the instrument does not get used for things it cannot do.

It does not tell you whether the game is good. It tells you whether it is a game. A build can score 7 and still be tedious, because tedium lives in tuning, and tuning is not on the sheet.

It does not survive contact with a broken loop. A high score with logic that misfires is still broken. That is a different diagnosis with a different procedure — locate the state that disagrees with the screen, then patch the rule, not the symptom.

It is not a market signal. Nothing here says whether anyone wants this. It only says whether what you have is a game, which is a precondition for the question, not an answer to it.

It should not be re-run after polish. Re-grading a build you have spent six hours polishing produces a higher score for reasons that have nothing to do with the axes, because you are now emotionally invested in the answer. Grade once, before the build; confirm once, after. Then stop measuring and go tune.

The one-page verdict sheet

Copy this, fill it, keep it next to the prompt box. The upper half is the prediction; the lower half is the confirmation.

PREDICTION (before the first prompt)

idea in one sentence: ______________________________
the verb, one word:  ______________________________
the number:          ______________________________
the price of failing:______________________________
the second run:      ______________________________

A1 ___  A2 ___  A3 ___  A4 ___     TOTAL ___ / 8

band 0-1 scene / 2-3 demo / 4-5 convertible / 6-7 game / 8 game
if under 6, the zero axis is: ______  ladder rung: ______


CONFIRMATION (after the build, stranger, no narration)

stopped acting deliberately at: ____ s
stopped entirely at:            ____ s
said "again" at:                ____ s
failed signals (list numbers):  ______________________

verdict:  game / convertible demo / demo / scene

Two lines decide most of it: the zero axis in the upper half, and the "again" number in the lower half. If the zero axis is A2, you are one sentence away from a different build. If the "again" number does not exist, no sentence will save this one, and the honest move is to write the next pitch rather than the next prompt.