AI Vibe Coding: Faster Doesn’t Mean Better

Setting the Stage

In the late 90s, Alan Cooper wrote The Inmates Are Running the Asylum. His warning was simple: when engineers design for possibility instead of humanity, we get software that works technically but fails the people who depend on it. We’ve lived with that trade-off for decades, but the pace of change gave us room to self-correct. Frameworks matured slowly. Communities learned from painful mistakes. Progress wasn’t perfect, but it bent toward usability over time.

---

The Core Systemic Issue

Progress takes time. Not just calendar time, but the cycles of feedback, failure, and correction that shape a foundation into something sturdy. Every framework—Rails, Angular, Kubernetes—was generally fit for purpose when created; it learned from the projects built on it and evolved as technology, regulations, and expectations changed. That learning loop was possible because adoption, while rapid, wasn’t instantaneous.

Now enter AI-driven development. What once meant 100 apps in a year could soon mean 10,000. The acceleration doesn’t just change the curve of productivity—it compresses the feedback loop until it risks disappearing. If a once-fit-for-purpose framework becomes the substrate for thousands of AI-generated apps overnight, when exactly do we get the chance to notice and correct the cracks?

---

The House of Cards at Scale

We’ve always built on imperfect foundations. That’s not negligence—it’s how progress works. But historically, the slower pace allowed us to pause, patch, and iterate. AI threatens to turn this iterative process into a runaway train. Instead of a thousand missteps nudging us toward better practices, we could find ourselves locked into brittle patterns, scaled across industries before anyone can intervene.

Frameworks need to stumble to learn. Without that stumbling, there’s no growth. When AI strips away the time to stumble, it also strips away the opportunity to progress.

---

What Needs to Change

If AI is going to accelerate development, it must also accelerate correction. That means:

• Validation baked in: AI outputs must pass machine-checkable gates before they can be trusted.
• Design honesty: models should surface caveats and risks, not bury them under confident prose.
• Slower is sometimes faster: in critical systems, enforce deliberate pauses to allow human review and systemic learning.

Non‑Functional Drivers for AI‑Accelerated Builds

Before generation, decide what kind of thing you’re building. Non‑functional requirements (NFRs) should steer the design as much as features:

Expected lifespan

• Disposable (≤ 4 weeks): hack-fast, minimal scaffolding, ephemeral storage, explicit end-of-life date.
• Seasonal/ campaign (1–12 months): light telemetry, automated tests on the critical path, migration plan for data, documented decommission steps.
• Sustained/ core (12+ months): full traceability, versioned contracts, SLOs/alerts, upgrade paths, rollbacks, and security reviews.

Support & SLOs

Define response/repair targets (SLA/SLO). An SLA is the contractual promise to customers, while an SLO is the engineering objective (e.g., 99.9% uptime) that helps ensure the SLA is met.. If someone will page you at 2am, design for operability: health checks, backpressure, circuit breakers, and safe-to-retry semantics.

Traceability

Documentation is often the first victim of the schedule. For years we got away with incomplete documentation because we spent months and years building the code, becoming intimately familiar with it. If humans are no longer immersing themselves in the code, then we will need much more comprehensive documentation.

• Requirements ledger: who asked for what, why, and when.
• Decision log: key trade-offs with date, context, and alternatives.
• Dependency manifest: libraries, models, prompts, datasets—each pinned and checksummed.
• Provenance: model + prompt versions stamped into artifacts; reproducible runs.

Observability by default

Structured logs, metrics, and traces with correlation IDs. Dashboards created with the first commit, not the last.

Governance & data hygiene

Data classification, retention/erasure, PII handling, consent, model-use restrictions, and auditability baked in.

Backout and exit strategy

Every system—AI or not—needs a way to step back or shut down safely. This includes a backout plan (rollback procedures, data migration reversals, configuration restores) and an exit plan (kill switch, feature flags, and decommission playbook for data export, URL tombstones, and user communications).

Testing gates match the NFRs

• Disposable: lint + smoke tests.
• Seasonal: plus contract tests and basic load tests.
• Sustained: plus property tests, chaos drills, backup/restore rehearsal, failover.

---

Why This Matters to AI Builds

When code is generated quickly, humans don’t accrue the tacit knowledge that usually comes from long development. Traceability and observability substitute for that lost lived experience. Time still teaches—but when we compress the time, we must encode the lessons (requirements, dependencies, decisions) into the system itself.

---

Closing Thought

The promise of AI vibe coding is intoxicating: apps and services an order of magnitude faster than before. But faster doesn’t mean better. If we eliminate the time it takes to correct our mistakes, we also eliminate the progress that mistakes make possible.