The Best Demo You Should Never Ship

OpenClaw is the best example we have of vibe coding working and why you should never ship vibe code at the same time.

It does what vibe coding promises. You get rapid progress with almost no ceremony. Features appear quickly. The system feels alive. You interact with it and it responds in ways that make it feel coherent and capable. As a demo, it is compelling enough that people start embracing it openly as a tool rather than an experiment.

That is the seduction. Here is the truth.

What Reliance Reveals

The system is unstable in ways that matter once reliance enters the picture. Bugs are consistent. Security issues are obvious and unresolved. Sessions go silent when they need to stay active. The TUI can feel responsive while quietly queuing prompts, then execute actions late or out of sequence. After prolonged usage the TUI utterly fails. UI streaming stops mid response and only resumes after a refresh or restart.

I can list more. The point is not the count. Nothing breaks decisively. It just degrades over time.

Vibe code does not fail. It rots.

This is both a failure of execution and the nature of the approach. Vibe coding optimises for momentum and output, not for stability. The system grows by accretion. When something works, it stays. When something misbehaves, it is often worked around rather than understood. Over time, you end up with behaviour that functions just well enough to continue, but not well enough to stand behind. Not well enough to reason with. You produce slop.

The Design Tells the Same Story

That pattern shows up immediately in design decisions. On a good day the slop stays confined to the code. On most days it bleeds into everything the user sees.

AI can generate a panel layout quickly. It cannot tell you why one curve is wrong relative to the rest of the interface. It does not see the break in visual rhythm unless you already see it and force the issue repeatedly. At that point the correction is not coming from understanding. It is coming from constraint. You are the guardrail, not the model.

Colour exposes this even more clearly. AI can select a shade of gray that satisfies the prompt. It cannot tell you why that gray fails in sunlight on a mobile device, because you did not encode every constraint that experience has taught you to check. Contrast in glare. Device variance. Environment. Those lessons can be introduced into the prompt, sure. But when you have a veteran designer in the mix, you do not need to remind them. They carry scar tissue. They know to check before they ship because they shipped without checking once and it cost them.

There is no pause to ask why something is shaped the way it is. Only whether it appears to work. Inexperienced people cannot see the gap. Experienced people stop caring because the velocity feels like progress. Both paths converge on the same outcome: an artefact that looks finished and is not.

The Attack Vector You Volunteered For

When it finally misbehaves in a way that matters, it still goes unnoticed. Because the system trained everyone around it to accept small failures as normal. The degradation was gradual enough that nobody recalibrated their expectations. The baseline drifted.

OpenClaw right now is the most laughable attack vector you could possibly install on your machine.

Every unresolved security issue in a vibe coded tool is a feature request written by an attacker.

// The uncomfortable audit

That is not hyperbole. The security posture of a system built by accretion, where bugs are worked around instead of understood, where sessions silently fail and prompts execute out of order, is not a system with vulnerabilities. It is a vulnerability with a UI.

The demo was excellent. The tool is dangerous. The distance between those two statements is exactly the gap that vibe coding cannot close, because closing it requires the kind of slow, deliberate, unglamorous work that the approach was designed to skip.