A Developer Let Claude Debloat His TV — Here's Why You Shouldn't

A developer gave Anthropic's Claude agent full debugging access to his four-year-old Android TV, ordered it to disable unused apps, and now reports the television runs better than it did when new. Mert Cobanov published his entire prompt workflow on X and encouraged followers to replicate it. That enthusiasm deserves scrutiny. Handing an AI agent system-level control over a consumer device carries real risks — instability, disabled critical components, broken system updates, and exposure to prompt injection — and most of what Claude did can be done faster and more safely by hand.

The story matters because it sits at the intersection of two trends. Smart TVs are accumulating more pre-installed software than ever, and AI agents are gaining the ability to act directly on operating systems rather than just answer questions. When those trends collide on a device most people don't think of as a computer, the failure modes are worth understanding before they show up in your living room.

TVs are full of software you never asked for

Start with the scale of the problem. By the end of 2026, more than 50 percent of households worldwide are expected to own a smart TV, up from 34 percent in 2020. Almost none of them arrive clean. Samsung TVs, for example, ship with roughly 20 percent of their total storage already consumed by the operating system and bundled software.

Much of that pre-loaded software consists of streaming apps — Netflix, YouTube, Prime Video — and some manufacturers prevent customers from uninstalling them at all. The rest is often worse: software built to collect data about user behavior for advertising purposes. Automatic Content Recognition (ACR) regularly captures screenshots of whatever is playing on the television. The Center for Digital Democracy has called connected TVs a "privacy nightmare" for viewers. That same software can also slow down the operating system, which is what pushes frustrated owners toward drastic fixes.

What Cobanov actually did

Cobanov opened the developer options on his Android TV, then instructed Claude to connect and debloat the device. His published prompt included some guard rails: the agent was told not to uninstall anything, only to disable unused applications, shorten animation durations, and keep a record of every action for later review. Those constraints matter. Disable is reversible in a way that uninstall is not.

He went further than the prompt, though. Cobanov also disabled Google TV's home screen, which he said is largely filled with ads and "recommended for you" content. In its place he installed FLauncher, a free and open-source app available on the Play Store that displays a simple grid of the TV's installed applications.

The results, by his account, were significant.

"The big payoff was that the TV now runs smoother than when it was new, four years ago," Cobanov posted on X. He added that the entire operation ran without root access, meaning Claude never gained the deepest level of system control available on Android.

Why this is a bad template to follow

Cobanov's guard rails were reasonable. They do not make the workflow safe for everyone.

The first problem is the model itself. AI agents make mistakes, and they are less likely to admit when they don't know an answer. If Claude hallucinates a package name or misjudges which application is safe to disable, it can turn off a component the system needs. The consequence ranges from an unstable interface to a TV that won't boot properly.

The second problem is permanence. If key packages end up uninstalled or disabled, the television may be unable to receive system updates in the future. A TV cut off from updates is a TV with unpatched security holes — a particular concern on a platform already described by the Center for Digital Democracy as a privacy nightmare.

The third problem is external. Prompts can be hacked by malicious actors. An agent with debugging access to your TV, following instructions pulled from an environment an attacker can influence, is a meaningful attack surface on a device that sits inside your home and, thanks to ACR-style software, already knows a lot about your viewing habits.

None of these risks materialized for Cobanov. That is a sample size of one, and a developer at that — someone with the background to diagnose and repair a broken TV if the agent went sideways. Most owners do not have that fallback.

The manual route is shorter than you think

Here is the part that undercuts the whole exercise: nearly everything Claude did to this TV is possible — and safer — for a human to do manually.

FLauncher, the app Cobanov used to replace the ad-heavy Google home screen, is available for download on the Play Store without granting any AI system access. Other apps in the Play Store perform similar functions, letting you choose which packages run on your TV and which don't.

Sometimes the fix is even simpler. Freeing up memory can be as easy as clearing the cache, with no app deletion or disabling needed at all.

For performance specifically, Android TV's settings app contains a Developer Options section with controls that address the same slowdowns Cobanov targeted. Once developer options are enabled, you can reduce animation scales from 1x to 0.5x, which makes the interface feel faster and more responsive because the system spends less time drawing transitions. You can also disable autoplaying videos on the home screen, turn off usage diagnostics, and enable "apps only" mode to strip the interface down to what you actually use.

Each of those steps takes a few taps. None of them requires an agent with system access, and none of them can be derailed by a hallucination.

The bigger picture

Cobanov's experiment is a preview of a debate the industry will have repeatedly over the next few years. Agent frameworks are gaining the ability to execute multi-step operations on real systems, and consumers will keep being tempted to point them at tedious chores — debloating a TV, pruning a phone, cleaning a laptop. The advice from this episode is straightforward: let the agent suggest, and let the human execute. Publish the list of packages worth disabling, by all means. But the button that disables them should stay under a human finger, especially on hardware that manufacturers make difficult to restore.

For smart TV owners, the practical takeaway is that the bloatware problem has existed for years and has accumulated manual solutions — launchers, animation tweaks, cache clears, apps-only mode — that carry none of the agent risk. As agents grow more capable, the question for each of these chores will be whether automation adds anything beyond what a checklist already delivers. For this one, it doesn't.