What Is This?
Vibe coding — throwing casual prompts at AI and watching code materialise — feels like magic. The productivity high is immediate and real. The problem shows up three weeks later, when you've burned $300 in API costs, have a graveyard of 85%-complete projects, and a codebase that nobody (including the AI) can maintain.
@Eli5defi's "Confessions of a Vibe Coder" is a practitioner's honest post-mortem: the author went from $300/month with zero shipped features to $150/month with 8-12 shipped features per month by applying three core disciplines: the 85% Rule, model arbitrage, and structured prompting.^1
The verdict isn't "AI coding doesn't work." It's that AI is a tool that demands the same discipline you'd apply to managing a junior developer — scope definition, context handoff, task matching, completion forcing — and without that discipline, you're paying a lot to generate expensive scaffolding you'll never ship.
Why Does It Matter?
- The 85% trap is real and expensive. Every abandoned half-built feature represents 10-20 hours of prompting work and the API costs that went with it. The sunk cost isn't just time — it's momentum. Vibe coders accumulate a graveyard of almost-products.
- Model arbitrage can cut your AI costs by 60-70%. Using Claude Opus for everything is like using a scalpel to dig a ditch. Opus for planning, Sonnet for implementation, Haiku for debugging — the math works out to $5.50 vs $15+ for the same output.^1
- Prompt caching changes the economics completely. After the first request, your system prompt, codebase context, and documentation are cached. Subsequent requests see 90% cost reduction on those tokens. For a 50-iteration project, this is the difference between $50 and $5 in context costs.^2
- AI has no memory between sessions — you have to create it artificially. Without an agent journal (a persistent file where the AI logs progress, findings, and build states), every new session restarts from zero. You're not managing a developer; you're repeatedly hiring one who doesn't remember the last three weeks.
- Spec-Driven Development (SDD) is the emerging discipline that fixes this. Rather than prompting toward an undefined target, SDD starts with a written specification — architecture decisions, success criteria, edge cases — before any code is generated. It's what engineers always did; AI just makes it even more necessary.^3
Key People & Players
- Andrej Karpathy — Introduced "vibe coding" as a concept in early 2025: accepting AI-generated code without closely reviewing its structure, relying on results and follow-up prompts to guide changes. Descriptive, not prescriptive — he was naming what he observed, not recommending it.^4
- @Eli5defi — DeFi educator and builder, "Confessions of a Vibe Coder" is the practitioner's honest account of what unstructured AI coding actually costs and what fixes it.^1
- Anthropic — The model pricing and caching architecture that makes model arbitrage possible. Prompt caching (90% discount on cached tokens) and the Opus/Sonnet/Haiku tiering were designed to enable exactly this kind of layered approach.
- Thoughtworks — Recognised Spec-Driven Development as one of the most significant engineering practices to emerge from the AI coding era in their 2025 Technology Radar.^3
The Current State
Vibe coding has become the default entry point for non-engineers building with AI. The term has gone from Karpathy's offhand description to a cultural phenomenon — YouTube tutorials, Reddit threads, X threads all celebrate the "just describe it and watch it build" experience. That framing captures the beginning of the experience, not the middle or end.
The backlash and correction are now emerging. TechRxiv published a formal paper on "Technical Debt-Aware Prompting Frameworks for Sustainable Vibe Coding" in late 2025.^3 Spec-Driven Development is gaining traction as the engineering discipline that tames the chaos. The pattern mirrors every previous tool revolution: early enthusiasm → painful reckoning with limitations → systematic methodologies → mature practice.
We're now at the reckoning phase. The builders who learn the discipline now will be the ones who actually ship.
Best Resources to Learn More
- @Eli5defi's original article — The practitioner's account that sparked this piece. Specific, honest, actionable.^1
- Softr's "8 Vibe Coding Best Practices" (2026) — Practical checklist: define intent, wireframe first, structure prompts in steps, document decisions.^5
- TechRxiv: "A Technical Debt-Aware Prompting Framework" — Academic treatment of the production readiness crisis in AI-assisted development.^3
- Anthropic API Pricing docs — Understand the model tiers and prompt caching before you optimise.^2
- "From Vibe-Coding Chaos to Spec-Driven Precision" (01cloud Engineering Blog) — The best practical overview of SDD as the systematic alternative.^3