AI is a Goldfish!' Never Trust AI When It Says 'I'll Do It All!

This article is a practical survival guide for those who have already introduced AI agent development environments like Antigravity and have started creating games and works alongside AI.

When you tell an AI, "I want to make a game like this!", it reliably replies, "Understood completely! I'll do it all!" Seeing it create files and write programs and scenarios at breakneck speed truly feels like magic.

However, when you seriously start creating a work with AI on the frontlines, you quickly realize something. You must never take the AI's "I'll do it all" at face value.

Why? Because AI has the memory of a goldfish.

1. AI Itself Doesn't Understand Its "Poor Memory"

The biggest problem when working with AI is that the AI itself doesn't understand the limits of its own memory (context size).

Because AI mistakenly believes its brain capacity is infinite, when a human asks it for a huge favor like "Build this whole feature perfectly!", it readily agrees, saying "I'll do it all!" without considering its context retention ability at all.

In reality, however, the amount of information an AI can retain at once (context window or token limit) is very poor.

Of course, current AI agents are somewhat smart, so they try to maneuver by saying, "I'll make a plan and divide the tasks."
But as the project grows, the information the AI "must reference to maintain consistency"—past scenarios, character settings, complex programs—explodes in volume.

This is why an initial "zero-to-one generation" looks great, but the moment you enter the phase of "modifying" existing code, the AI's performance appears to plummet.

When adding or modifying things, the context size required to maintain system-wide consistency leaves the AI's brain gasping for air, quickly causing it to exceed capacity.
An AI that has reached its limit will forget the rules from just a few steps ago, prioritize only the task right in front of it, and rapidly go out of control.

2. Humans Must Trim the AI's Grand Plans

So how do we prevent this out-of-control behavior caused by exceeded capacity?

The answer is for the human to mercilessly take a red pen to the Implementation Plan proposed by the AI, stripping it down to minimize tasks.

Before working, an AI agent will ask, "I'll proceed with this plan, is that okay?" At this point, the AI will casually present an unreasonable plan like, "I'll finish 10 bug fixes and feature additions all at once in this single run!"

If you believe this and give the green light, a mountain of bugs will be built a few minutes later.
That's why the human must intervene here and say, "Don't get greedy. Your brain can't handle that much," or "Just focus on this one thing this time," effectively trimming (detuning) the work scope.

The fundamental premise to keep a project from collapsing is for the human to understand the AI's "memory limits" and constantly keep its head clear by strictly enforcing "only one task per run."

3. Absolutely Prevent the Dreaded "Time Leap"

Even so, AI occasionally makes tremendous mistakes. It might accidentally delete an important program or break the game entirely.

The most terrifying thing here is when a panicked AI decides, "Oh no! I made a mistake, so I'll just rewind time to a slightly older dataset!" and arbitrarily tries to revert to past data (from a few hours or days ago when things were normal).

While this might be a correct recovery method systematically, for a creator, this is the "worst possible action" that you must never let it take.

Why? Because if time is rewound, all the "god-tier dialogue" or "best ideas" you squeezed out with the AI in the few minutes right before things broke will vanish as if they never existed.

I don't want to go back 3 days, or even 3 minutes. We need to fix the data in the "present," not run away to the "past."

4. One-Click Command Recovery is Banned! Force Manual Fixes in the Editor

So, when data gets corrupted, how do we fix it without rewinding time?

AI will often try to smartly fix things using terminal commands, saying, "I've replaced it with a clean file for you in one shot!" but this is also banned.

Recovery work must always be done by having the AI rewrite text on the "editor screen" that we are normally looking at.

Here's how to do it:
1. First, prepare the "clean file" from before things went wrong as a base.
2. Add the lost "best recent results" to that base. Even if the file disappears from the screen, everything the AI wrote just a moment ago is entirely preserved in the "chat history" right in front of you. Pick up that text from the chat and merge it into the clean base.

💡 The Core: Maintain the Power to Press "Ctrl+Z" at Any Time

Why make it rewrite things on the editor screen instead of using a single background command?

It's so that the human can press the "Ctrl+Z" (Undo) shortcut key at any given moment.

If a file is entirely replaced in the background without permission, the human loses the ability to say "Actually, never mind that edit!" and undo it. By making the AI work little by little on the editor, even if it starts making weird fixes again, you can immediately put a stop to it with "Ctrl+Z."

Conclusion: There Is No 100% Perfect AI

There is nothing less trustworthy than an AI confidently declaring, "It's 100% fine!"

The correct approach is to interact with them under the premise that "AI has the memory of a goldfish" and "AI's context capacity is poor."

Instead of leaving everything to the AI and trying to have it build things automatically, the human must continuously hold the reins in a messy, hands-on way: "Don't rewind time without permission," and "Work right in front of me so I can undo it anytime."

This is the most reliable survival guide for making it to the very end of the grand adventure of game development with AI.

💡 P.S.: The Ultimate Dilemma

To those who read this far and thought, "I see! Then to make sure the AI never makes a mistake, I just need to write a 'Supreme Rulebook' with 100 prohibitions and rules, and make it read that before working!"

If you do that, the AI's brain capacity (context) will become overloaded just trying to "remember the rules," causing its core task-processing ability to drop significantly, and ironically resulting in it producing far worse mistakes (laughs).

"We must bind the AI with rules," but "Making the rules too long makes the AI stupid."

Amidst this endless dilemma, I am personally still testing "how much automation can be left to AI," but one of the surest solutions at present is "proceeding one step at a time through 'micro-steps' with humans inserting detailed checks."

Well, I'm an IT amateur so automation is difficult for me, and since I'm glued to the front of my PC during work anyway, these "micro-steps" might just be enough for me (laughs).

...Ah, there is another quick fix: an approach of "using the latest, high-spec model with a massively huge brain (context)." But as the project bloats, I feel like we'd ultimately just run back into the exact same dilemma of "the usable token limit still isn't nearly enough!"...