karenizer

karenizer is a web app that takes your ordinary grievance — the wrong order, the rude rep, the package that arrived looking like it was delivered by a trebuchet — and transforms it into a full-throated, theatrically escalating complaint letter. You pick how unhinged you want it: from “Politely Disappointed” (a calm adult) all the way to “PEAK KAREN” (an entity addressing the CEO of a coffee chain as though they personally wronged the founding fathers). It is a deeply silly project. I am not sorry.

The whole thing is a prompt engineering exercise dressed as a joke

There are six Karen levels. Each one is a distinct system prompt, hand-tuned until it hit the right note. Level 0 is professional and measured. Level 3 has the LLM demand a manager, threaten Yelp, and deploy strategic capital letters FOR EMPHASIS. Level 5 — Peak Karen — instructs the model to invoke “the spirit of every wronged customer in history” and implies the grievance will be discussed on national television.

Writing six system prompts that produce reliably different tones without bleeding into each other is actually non-trivial. Level 2 needs to feel annoyed-but-controlled; Level 3 needs the switch to have flipped but not melted. Getting that gradient right meant a lot of test complaints about cold soup. Science demands sacrifice.

The intensity scale isn’t cosmetic — it’s the product. The same “my pizza was cold” turns into a completely different artifact at each level, and watching the tone shift is what makes it funny. The UI has a slider with an emoji that updates as you drag, because if you’re building something this ridiculous, the feedback loop should be immediate.

The model does drama. The code does nothing interesting.

There’s no ranking, no scoring, no multi-stage pipeline here. This is one API call, which is actually the correct choice. The job is purely expressive: take a description and produce theatrical text. There’s no deterministic logic to offload, no place where code is more reliable than the model.

What the code does handle: building the user prompt from the form fields (company name, category, the specific grievance), picking the right system prompt from a dictionary keyed by level, and validating the request so nobody submits a 10,000-character rant that costs me real money. Pydantic handles the typing. FastAPI handles the routing. LiteLLM handles talking to whatever model is configured via environment variable — I ran on Claude during development and swapped to another provider for deployment with a one-line change. That abstraction earns its keep on a project like this where the model is literally the entire product.

Temperature is 0.85. High enough that two identical complaints generate meaningfully different letters; low enough that Level 5 doesn’t veer into actual gibberish. It took about four tries to land on that number, which I mention only because “tune your temperature” is advice that sounds obvious until you’re reading output that’s either too stiff or has started making up legal citations.

No framework, no database, no problem

The frontend is vanilla HTML, CSS, and about a hundred lines of JavaScript. No React. No build step. No node_modules folder to make the repo feel heavy. A form submits to /api/generate, a hidden div swaps in with the result, a button copies it to clipboard. That’s the whole interaction.

This was deliberate. When the entire product is one API call and one text box, a frontend framework is a solution looking for a problem. The constraint of “no framework” also made deployment on Vercel’s Python runtime almost trivially easy — there’s no build pipeline to configure, just a FastAPI app with an api/index.py entry point and a twelve-line vercel.json.

The stack

Python 3.12, FastAPI for the API and Jinja2 templating, LiteLLM for model routing, Pydantic v2 for request validation, and plain HTML/CSS/JS for the UI. Deployed on Vercel. Three runtime dependencies. No database — there’s nothing to persist; the letter exists, you copy it, it’s gone. I like systems with no state to corrupt.

If you’ve built something equally useless but technically honest, I’d enjoy hearing about it — reach out.