Skip to content

AI feature bloat: why shipping less is better

November 17, 2026

5 min read

An engraved hand lifts excess feature cards away from a focused app with a single orange core.

Adding a feature used to cost a developer nights and weekends, so every app was a set of careful choices about what deserved the time. Now an agent will build whatever you ask it to, and you can ship an app on Saturday and bolt ten features onto it by Wednesday.

Hassan El-Mghari, director of developer experience at Together AI, has been building and launching apps since 2020, and he made the case against doing that on a recent episode of Rabbits on the Run.

Hassan’s process comes down to building the things people want and holding every one of them to a high bar, instead of adding features just because he can.

It’s very easy to ship an app and then add 10 features to it. But just because you can doesn’t mean you should. Shipping less is better.

Every feature you ship is a trust bet

Hassan describes being asked to try an app and not understanding its purpose because it does everything.

When I try to use it, like half the features are broken, and you just lose a lot of trust in that person.

Feature bloat feels like a design problem to the team building it, but users read it as a trust problem. An app with one feature that works tells people you care about the product and the user. An app with ten features, and half of them are broken, tells people you prompted until the demo looked good. Hassan has a name for that type of app:

That’s where things like AI slop come into play, where they’re like, what is this vibecoded mess? You don’t want to get to a stage where someone calls your app a vibecoded mess.

Everyone is shipping massive amounts of code now, so the details are what differentiate builders. When generation is abundant, polish becomes the scarce signal, and every half-working feature you ship spends trust you’ll want back later.

Review the idea before the agent builds it

Hassan tests interest before committing to a full build. He teases apps on social media before building them in earnest, and treats the response as a gate.

I’ve teased a lot of apps that I haven’t even released because the teaser didn’t go well. The teaser is a very good indication for how the app will actually perform when I launch it.

The cost of building used to force this kind of validation, since nobody hand-built a CLI for a market they hadn’t checked. Now that an agent does the work, the validation has to be deliberate, and Hassan has moved it ahead of the build.

The teaser also writes the roadmap. When five different people comment asking whether the app will have the same feature, that feature jumps the queue, and the launch gets framed around it. Scope stops being driven by whatever the agent produced in the session and is led instead by what people say they want in the app.

Launch in public and listen

Hassan launches everything in public, and the public is not shy with their feedback.

Sometimes I get some really harsh comments, but they are also right. Like, you have this big security vulnerability. This API route’s not even authenticated. What are you doing?

He treats this as a free review from people who took the time to try his work, and he’s right that it’s valuable. The unauthenticated route Hassan missed shows why independent review matters, regardless of whether the author is a person, an agent, or a person driving an agent.

Keep the scope to five words

Hassan’s practical bar for scope is almost comically simple.

Make sure you can describe your idea in five words. Just because you can ship all this stuff doesn’t mean you should.

The five-word limit keeps the focus on building one thing and making sure it works well. His most recent launch, an app that adds AI-generated subtitles to any video, shows the whole system running end to end. He saw a real problem, teased the app, read the comments to learn which features mattered, built those before launch, and shipped something you can describe in a breath.

The apps that fail the test, do so before the first prompt. If the description needs a paragraph, the app will need ten features, and ten features at agent speed means ten surfaces where trust can break. For the builders shipping this way, the review habits for vibe coding are worth picking up early.

Watch the full conversation

The episode goes well beyond scope discipline, covering how Hassan built an audience by learning in public, why he treats harsh commenters like customers, and what six years of launching apps taught him about reading signal from a timeline.

Watch the full episode of Rabbits on the Run

And if the unauthenticated-route story hit close to home, remember that an independent review can give you that feedback before merge, rather than on launch day. CodeRabbit reviews every PR with full codebase context, so the security flaw turns up in the pull request instead of the quote tweets.

Start a free trial

About the author

Share

Share on RedditShare on XShare on LinkedIn

Catch the latest, right in your inbox.

Add us to your feed.

GetStarted in2 clicks.