After Getting Burned by AI Coding for Three Months, I Built WishGraph

When Vibe Coding Hits the Wall of Reality

Over the past three months, a Bilibili creator ran a rather extreme AI coding experiment—and the result was getting "badly burned." Not only did it cost a fortune, it also wasted an enormous amount of time. This isn't a joke; it's the shared predicament faced by countless people trying to rapidly build projects with AI today.

Let's start with the bill: over 300 RMB spent on DeepSeek API fees (and that's the cheapest option). ChatGPT, even with the half-price Turkey region hack, still required monthly renewals. Claude Code and Google subscriptions each ran over $20 a month. To "use it more scientifically," the creator even set up a personal server—yet still couldn't escape the fate of getting the Claude Code account banned.

Money is just the surface. The real problem is this: the creator built one small web page after another, growable knowledge bases, apps, auto-formatting assistants... nearly every AI coding project imaginable was tried, only to discover that "the more you build, the more mistakes you make."

Why AI Coding Gets Messier the More You Do It

This touches on the most central contradiction in current AI-assisted development.

"The model is powerful, which means its generation speed keeps increasing, but the speed at which humans understand and verify hasn't kept pace."

This sentence precisely captures the crux of the issue. When a developer has zero understanding of a project's architecture, modules, and dependencies, the faster AI generates code, the faster technical debt accumulates—and the harder the black box becomes to control. The human role as the "verifier" gets left far behind by the high-speed generation capability.

Relying solely on prompts casually typed into a dialog box (colloquially known as "prescribing meds through the window") is far from enough—these instructions are vague, mutable, and hard to reuse. What's truly needed is to translate these "vague and mutable ideas" into detailed task specifications, and to distill them into a set of project rules that can be updated in real time. This is precisely the core problem WishGraph attempts to solve.

WishGraph: Splitting AI Into Two Windows

The creator ultimately stopped piling on projects and instead built a skill called WishGraph. Its core design philosophy is: completely separate discussion from execution.

The Discussion Window Handles Specs, the Execution Window Does the Work

WishGraph splits the AI workflow into two independent windows:

• Discussion window: Here you confirm requirements with the AI, and the AI produces detailed task specifications. Essentially, you let the AI help you write complete prompts first, clearly telling the executor—which files to read, what can and cannot be changed, and the final acceptance criteria.
• Execution window: Has independent context and clear instructions, focused on completing a single, well-defined task.

This separation brings two direct benefits. First, the execution action is isolated—while the AI works in the background, you can continue advancing new lines of thought in the discussion window, so efficiency is no longer bottlenecked by serial blocking. Second, the execution window no longer "guesses" your preferences from lengthy context but acts on clear instructions, dramatically reducing the chance of going off track.

Independent Execution Based on Git Clones

WishGraph adapts to the native capabilities of CodeX and Claude Code, allowing you to directly create sub-agents with independent context. This independent window typically runs on top of a Git clone. Once the task is done, you simply return to the discussion window and tell it to "inherit the results."

Even more interesting, you can even hand execution tasks off to other agents. In the video, the author gave a task to Claude Code, and by simply saying "execute task 026," it automatically located the corresponding task specification precisely and completed it.

Multi-Agent Collaboration and Task Integration

The true value of WishGraph lies in how it restructures chaotic single-window development into a controllable multi-agent collaboration workflow.

As long as there's no overlap between multiple execution windows, they can in principle run in parallel. The video demonstrates tasks 025 and 026 completing simultaneously; then, back in the discussion window, a single command—"integrate everything for me"—makes the system uniformly incorporate the task specifications and execution result reports into the integration, while also saving a state file that records the latest project progress.

This way, complex project development shifts from the old model of "running everything in one window" to "one person in the loop providing direction and judgment."

"We haven't eliminated AI and the project's complex black box, but we've improved its input, output, and control boundaries, left clear evidence, and helped ourselves better identify exactly where problems occurred."

This is a restructuring of the causal chain. The project itself hasn't become smaller or simpler, but it has once again become manageable. Interested developers can search for WishGraph on GitHub.

Epilogue: Will All Engineering Design Eventually Be Swallowed by Platforms?

At the end, the author offers a fairly clear-eyed judgment: this kind of engineering-level redesign will eventually be swallowed by the capability development of platforms like CodeX and Claude Code.

In other words, the dual-window collaboration framework painstakingly built today may well become a default feature of official products tomorrow. But he isn't discouraged by this—instead, he extrapolates a grander trend—

"It's not just this AI coding path that will be swallowed; it's very likely that our operating systems will gradually transform from Windows, iOS, and macOS into an Agent OS."

This so-called Agent OS means that in the future, when you enter an operating system, you'll no longer click on graphical icons one by one, but instead converse with an Agent to have it create a highly personalized way of working for you. The author believes this is an underlying trend—it just may take a very long time before it's proven out.

For developers today, WishGraph's takeaway may be more practical: facing the risk of AI coding spiraling out of control, the answer isn't to use it less, but to set clearer boundaries, more explicit specifications, and more controllable collaboration workflows for the AI. In an era of skyrocketing model capabilities, humans need to reclaim their position of "giving direction and making judgments."