I’m Going Back to Writing Code by Hand#

Original | Raw

Author: shvbsle (k10s.dev) Published: 2026-05-09

Summary#

A developer vibe-coded a GPU-aware Kubernetes TUI (k10s) for 7 months using claude-code, then archived the entire codebase and started over. The project grew from a focused GPU dashboard to a bloated k9s clone because AI-generated velocity made every feature feel “free.” The resulting 1,690-line god object with 110 switch/case branches became unmaintainable. The post extracts five tenets — hard-won lessons about what AI gets wrong when projects grow complex.

Core thesis: AI writes features, not architecture. The longer you let it drive without constraints, the worse the wreckage gets.

The Five Tenets#

Tenet 1: AI Builds Features, Not Architecture#

Each feature was implemented in isolation — “make this work right now” — without awareness of 49 other features sharing state. Nine manual = nil cleanup assignments scattered across the file. Miss one → ghost data from previous views.

Fix: Write architecture yourself before any code. Put invariants in CLAUDE.md:

  • Each view implements the View trait. Views do NOT access other views’ state.
  • Adding a new view MUST NOT require modifying existing views.
  • The App struct is a thin router — navigation and dispatch only.

Tenet 2: The God Object Is the Default AI Artifact#

AI gravitates toward single-struct-holds-everything because it satisfies the immediate prompt with minimal ceremony. Result: one keybinding (s) meant three different things depending on view state. 20+ occurrences of string-comparison type discrimination in one file.

Fix: State ownership rules in CLAUDE.md:

  • NEVER add fields to the App struct for view-specific state
  • Each view declares its own key bindings
  • If your change requires modifying existing views, stop and ask

Tenet 3: Velocity Illusion Widens Your Scope#

Vibe-coding makes everything feel cheap. The developer went from “GPU dashboard for cluster operators” to “general-purpose Kubernetes TUI for everyone” because each feature took one session. But: infinite line budget ≠ infinite complexity budget.

Fix: Explicit scope boundary in CLAUDE.md:

  • Define who you’re NOT building for
  • Reject features that don’t serve the core audience
  • Say no in advance, before the velocity high convinces you to say yes

Tenet 4: Positional Data Is a Time Bomb#

All Kubernetes resources were flattened to []string. Column identity was purely positional (ra[3] = Alloc). Add a column → every sort, conditional render, and accessor silently breaks. Compiler can’t help because it’s all strings.

Fix: Never flatten structured data into positional arrays. All data flows as typed structs until the render() call. Sort functions operate on named fields, never indices.

Tenet 5: AI Doesn’t Own State Transitions#

A closure mutated Model fields from inside a goroutine while View() read the same fields on the main goroutine. No lock. Textbook data race that worked 99% of the time.

Fix: Background tasks NEVER mutate UI state directly. They send typed messages through a channel. Only the main event loop applies mutations. render()/view() is a pure function.

Meta-Insight: CLAUDE.md as Architecture Enforcement#

Every tenet’s fix is the same pattern: write the constraint in CLAUDE.md so the AI sees it on every prompt. The AI will follow rules if you write them down — it just won’t invent them for you. The developer’s job is to make the correct path also the shortest path.

Connections#

Key Quotes#

“AI writes features, not architecture. The longer you let it drive without constraints, the worse the wreckage gets.”

“You have infinite LINE budget. But you have the same finite complexity budget as always.”

“The AI will follow these if you write them down. It just won’t invent them for you.”

See Also#

Backward Trusted Remote Execution Rex Activity Log Forward