Stop Vibe Coding: Why 'Think Slow, Act Fast' is the Only Way to Build Software in the AI Era

The GPT revolution has fueled a dangerous popular myth: that "vibe coding" - rushing immediate output using AI tools - is highly productive. Any seasoned developer knows the reality is a disaster in practice. Chasing immediate output, burning tokens, and letting context fill and hallucinate quickly leads to a spiraling set of errors and break-fixes that balloon technical debt.

The principle that rigorous planning must precede rushed execution is not new. I recently revisited the terrific book "How Big Things Get Done" by Flyvbjerg and Gardner. By studying massive national infrastructure projects, their key refrain for success is simple: "think slow, act fast."

Last week, I applied that exact mantra to a major refactoring exercise. The results were startling proof that this approach is the ultimate AI-era engineering strategy.

Phase 1: Think Slow (4-5 Days, Zero Lines of Code)

My focus was 100% on preparation, utilizing AI as a rigorous assistant, not a primary coder.

Specs and Collaboration: I spent 4-5 days, on and off, creating comprehensive specifications and documentation, including vital conversations with colleagues to gather ideas.

LLMs as Critics: I engaged various LLMs (Claude, Codex) to read and critique the specs. I then rewrote, reorganized, and clarified points, throwing out anything that didn't align with the core requirements.

No "AI Slop": Every single sentence generated by the LLMs was read - thoroughly - multiple times. I argued points, brought my engineering experience to bear, and insisted on fine modifications to ensure the documentation aligned perfectly with intended customer and operational needs.

The Grumpy Architect Test: I ran our own agents against the draft specifications. My favorite agent, "Grumpy Architect," ruthlessly kicked back on unnecessary cruft, citing YAGNI (You Ain't Gonna Need It) and DRY (Don't Repeat Yourself).

The result of this slow, deliberate phase was about 2,500 lines of spec written and added to the repo, forming a thorough set of requirements, acceptance criteria, architectural decisions, and experience-led arguments. Crucially: In all this time, not a single line of application code was written.

Phase 2: Act Fast (Less Than 30 Minutes)

When the planning was unequivocally complete, and I finally said "do it," the execution was nearly instantaneous.

The total agent coding time was less than 30 minutes. Bar a few minor tweaks, the refactoring was finished and functional shortly after that. This refactoring exercise proves that in the AI-powered engineering world, preparation is the ultimate accelerant, with AI providing the necessary force multiplier.

Don't let the ease of generating code tempt you into a costly cycle of spiraling technical debt. Embrace the power of thorough specification and use AI to challenge your assumptions and solidify your design - not just write your code.

Think slow, act fast.