I'm going back to writing code by hand
AI coding tools are enabling developers to generate large codebases quickly, but many report that “vibe-coding” without clear architecture or careful review leads to unmaintainable spaghetti code and hidden bugs. Commenters argue that current models are good at filling in features and boilerplate, not at making coherent long-term design decisions, so humans still need to own system architecture, constraints, and code review. There is broad support for using AI as an accelerated junior assistant within strong tests and specs, alongside concern that managerial pressure for speed and blind trust in agents will amplify technical debt rather than reduce it.
Scope of AI in Coding: Features vs. Architecture
- Many agree with the article’s core claim: current LLMs are good at implementing features but poor at making or evolving sound architecture.
- Common pattern: if you let agents “vibe-code” without strong constraints, you get god objects, tangled state, and hidden dependencies.
- Several argue architecture must be designed by a human first (interfaces, ownership, message types), then given to AI as a spec; AI should mostly fill in functions and boilerplate.
Code Quality, Testing, and Refactoring
- Strong emphasis that AI-generated code must be reviewed like a junior’s work; failure to read the output is seen as the real problem.
- Others counter that reviewing all AI output at scale is unrealistic, so test coverage (unit, integration, e2e) becomes the main safety net.
- Multiple comments report AI being particularly bad at large refactors: it tends to add code, miss nuances, and create subtle inconsistencies even under lint rules.
- Desire for a “verification layer” or better tooling to automatically check AI’s work beyond conventional tests.
Workflows and “Comprehension Debt”
- Many propose rules: only let AI write code you could write yourself, and don’t ship AI code you don’t understand; otherwise you accrue “comprehension/cognitive debt” that later becomes unmanageable.
- Some find AI most useful for: scaffolding, codemods, tests, exploratory design, or boring boilerplate (CRUD, forms), while keeping humans in charge of core logic and abstractions.
- Others experiment with strict modularization, small-scoped tasks, plan-first modes, and detailed skills/spec files to keep agents from derailing.
Productivity, Enjoyment, and Management Pressure
- Mixed experiences on productivity: some claim 50–100% coding time savings; others say perceived speed hides slower real progress and massive tech debt.
- A recurring concern: managers now expect “AI-speed” delivery and may push vibe-coded features, offloading cleanup onto developers.
- Several devs say they’re scaling back agentic coding for personal projects because it’s less enjoyable and produces sloppier code long-term.
Meta: Title, Authenticity, and Hype
- Many criticize the post title as clickbait, noting the author still uses AI for implementation while only “doing design by hand.”
- Some suspect the blog post itself might be LLM-written, and more broadly see the article as part of a broader, overhyped AI-marketing narrative.