7 months of vibe coding in exchange for 1,690 lines of god object: Developers abandon AI and rewrite manual code

Core conclusion

On May 12, a developer disclosed a complete review of his seven months of "vibe coding" (generating code at will with AI) on Claude - he spent 30 weekends and 234 submissions to build a GPU-aware Kubernetes TUI tool k10s, and finally got a 1,690-line God Object. He decided to archive the entire project and write it by hand from scratch.

This decision received 956 points and 581 comments on Hacker News, becoming one of the hottest technical topics of the day. It also represents the rising wave of "anti-vibe coding" reflection in the field of AI-assisted programming.

Key Points

• Event time: May 9, 2026 (review article released), topped the list of HN on May 12
• Scope of influence: All developer teams using AI coding tools, especially projects that rely on "vibe coding" (brainlessly letting AI write code)
• Core changes: The cognitive change from "AI writes code faster" to "AI writes code requires humans to design the architecture"
• HN heat value: 956 points, 581 comments, the hottest technology topic of the day

Background: From Magical Experience to God Object

Developer shvbsle launched the k10s project in September 2025 - a GPU-aware Kubernetes TUI dashboard inspired by k9s but designed specifically for NVIDIA GPU cluster management. He used Claude to generate code every weekend, and the initial experience was like magic:

"The first few weeks were pure magic. I told Claude 'Add a Pod view with live updates' and it worked. Resource lists, namespace filtering, log streaming, description panels, keyboard navigation - every feature was implemented cleanly because the project was small enough for the AI to hold the entire context."

The basic k9s clone took 3 weekends to complete. Pod/Node/Deployment/Service views, command palette, watch-based real-time updates, Vim shortcuts - all generated by Claude in a single session. His development speed increased by about 10 times.

The turning point came after adding the GPU cluster view. Claude generated a complete cluster view at once - beautiful tables, color-coded status bars, GPU/CPU/All filters - but when he entered the command to switch back to the Pod view, the table was blank. Live updates stopped. The node view shows the old data of the cluster view.

**God Object has been formed. **

Key Impact: 5 Costs of AI Coding

Core lesson: AI is good at writing functions, not designing architecture

The author distills 5 core teachings that have been verified by real projects:

1. AI builds functions, not architecture — Every time Claude is prompted to add a function, Claude delivers it perfectly. But each function is implemented in the context of "make it available now" and has no regard for the same state space shared by the other 49 functions. Solution: Write architectural rules in CLAUDE.md for the AI ​​to see at every prompt.

2. God Object is the default output of AI — AI prefers the "single structure contains everything" solution because there is zero ritual. But the lack of view isolation makes keyboard handling a nightmare - the same s key is "auto-scroll" in the log view and "shell" in the pod view.

3. Unconstrained speed is an illusion — The key finding cited by the author: 9 manual nil assignments in the code scattered throughout the 1,690-line file. Miss one, and "ghost data" from the previous view will contaminate the current view.

4. Humans must write the architecture — not a vague design document, but a concrete set of interfaces, message types, and ownership rules. Then put the rules into CLAUDE.md.

5. Reading code is to find problems, not to accept "it can be compiled" - In 7 months, he only looked at diff to verify compilation, and never read model.go completely.

Adaptation suggestions

If your team or project is using AI coding tools (Claude Code, GitHub Copilot, Cursor, etc.), the following measures can avoid repeating the mistakes of k10s:

• Create CLAUDE.md or AGENTS.md in the project root directory and write architecture invariants (each view implements the View trait, views do not access each other's status, and new views must not modify existing views)
• Do a complete code architecture review at least once a week, paying special attention to the status management part
• When the core model file of the project exceeds 500 lines, set the "downgrade from vibe coding" signal
• Let AI be responsible for function implementation, and humans be responsible for interface definition and architecture design
• Use AI code review plug-ins (such as adamsreview) to conduct structured reviews to make up for AI's architectural blind spots

Tool entry

The discussion in this article naturally involves a number of AI tools: Claude (Anthropic's AI coding assistant), OpenAI, GitHub Copilot and Hermes Agent are all AI Agent tools that can be used for automated coding. AI coding is moving from the "brainless generation" stage to the mature stage of "human design + AI implementation".

Reference link

• Original text: I'm going back to writing code by hand
• HN Discussion: 956
• k10s project (archived version): GitHub

Internal link guidance

• Want to learn how to use AI coding tools correctly? Recommend this tutorial: AI Agent-Driven Content Automation: n8n MCP Building Guide from Scratch - Learn how to use AI Agent in a structured way to avoid the God Object trap.
• Real case: See how independent developers avoid the vibe coding trap and use Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo to achieve monthly income $5,000’s practical experience.