An Open Letter to Claude and Its Audacity

A Harvard Crimson op-ed by Ruslana Vardanyan is making the rounds, and it is not kind to the current state of AI-assisted web publishing. Written as a letter addressed directly to Claude, it skewers the growing number of websites that were clearly built on a single prompt and never tested: mismatched fonts, submit buttons that vanish on click, and AI-generated restaurant menus full of visual gibberish. The piece is funny, but the complaint underneath it is real and worth taking seriously if you are shipping anything to a browser in 2025 or 2026.

What happened

Ruslana Vardanyan, writing in the Harvard Crimson’s Flyby blog, published an open letter to Claude, Anthropic’s AI assistant, taking aim at what she calls “half-baked nonsense cooked up in the Chat cauldron.” The piece is satirical, but the examples are specific enough to sting.

She describes landing on a website with four mismatched fonts and a submit button that, when clicked, simply disappeared. She also calls out AI-generated menus for small restaurants, describing the imagery as “gloop and glop riddled with worms,” a phrase anyone who has seen Midjourney food photography gone wrong will recognise instantly.

The piece also takes a swipe at Harvard’s own my.harvard portal, which she says allegedly was “vibecoded” (built by describing what you want to an AI rather than writing the underlying code). She notes that the old site still loads first before the new one replaces it with a delay, describing it as a “jumpscare.”

Why it matters

The Crimson piece is a cultural signal, not a technical one. When an undergraduate newspaper is publishing op-eds about broken AI websites, the pattern has become visible enough to annoy ordinary users, not just developers.

The specific failures she names are instructive. Mismatched fonts and broken form submissions are not design taste issues. They are quality-control failures. A form that does not submit means leads are being lost. A restaurant menu that looks machine-generated tells customers the business does not care about presentation, which is a problem when food is literally the product.

The piece also gestures at something more structural. She writes that “everyone is just gonna ride this train until the wheels fall off,” suggesting that the incentive to ship fast with AI tools is currently stronger than the incentive to ship something that works. That is accurate. The correction will come from customers refusing to tolerate it, not from toolmakers.

For businesses that are actively publishing AI-assisted content or using AI tools to build pages, this is a good moment to audit what is actually live. Our AI integration work always includes a human review layer precisely because models do not catch their own layout or logic errors.

Is vibe-coding actually that bad?

Not inherently. Using an AI to generate a first draft of a page, a component, or a form is a legitimate time-saver when someone with enough context reviews the output before publishing. The problem Vardanyan is describing is not AI-assisted development. It is AI-as-the-entire-process, with no human check at the end.

The submit button that vanished is a good example. That is a bug any 30-second manual test would catch. The four mismatched fonts are a design call any human eye would flag. Neither failure requires technical expertise to spot. They just require someone to actually click through the page before hitting publish.

She also makes a fair point about alternatives. A $50 website from a local high schooler or a basic Wix build may not be beautiful, but it probably works. A non-functional page built on a four-sentence prompt is objectively worse, regardless of how quickly it was produced.

If you are unsure whether your current site is holding up, a care plan with regular audits is a lower-cost fix than discovering the problem through a spike in bounce rates or a missed lead.

Our take

Vardanyan’s piece is a student opinion column, not a technical post-mortem. But it captures something we see in actual client briefs: a growing number of business owners who published an AI-generated page, assumed it worked, and are now quietly wondering why their form submissions dropped off or why customers are commenting on how odd the site looks.

The real issue is not Claude or any other model. It is the absence of a review step. No AI tool ships with a “is this actually usable” checker baked in. That part is still on the human. A broken submit button is not an AI failure. It is a process failure.

We are also a little skeptical of the nostalgia angle in the piece. A Wix site from 2014 was not better because it was human-made. It was better because whoever built it tested it. The same standard applies today, regardless of what tool generated the underlying code.

If you are building with AI tools and want a second set of eyes before anything goes live, that is exactly the kind of work we cover in our web development engagements. And if you are an established business worried your site is quietly failing visitors, our news coverage of AI and web tools is a reasonable place to stay current on what is shifting.

What to do about it

1. Click every button and fill every form on your site today. Do it in Chrome, Safari, and on a phone.
2. Check your font stack. Open the page and count how many typefaces are rendering. More than two is usually a sign something went wrong.
3. Pull your form submission data from the last 30 days and compare it to the same period before your last site change.
4. If you used an AI tool to generate any page or section, ask a person who had nothing to do with building it to try to use it.
5. If issues keep appearing after fixes, consider a structured audit rather than patching one bug at a time.

The simplest quality bar: if a first-time visitor cannot complete the one task your page is designed for, it does not matter how fast it was built.

Source: Bing News · Claude AI