Hi everyone! 👋 I started building Pseudo while getting into LeetCode. Algorithms have always been compared by time and space complexity, but AI has become part of many developers' workflow. I wondered whether there should also be a way to measure how much AI was actually used. Token count didn't seem like a fair comparison across models, so I started experimenting with a metric I call Resource Units (RU). At the same time, I ran into another problem. Whenever I got stuck and asked an LLM for help, it would often steer my incorrect approach toward the correct solution. That made practice feel less like problem solving and more like memorizing someone else's algorithm. I wanted a tool that would only implement my reasoning, not improve it. Pseudo is the result. You describe your approach in natural language, and it generates code in your chosen language using your chosen model. The system prompt is intentionally constrained to implement exactly what you describe, right or wrong. If your approach is flawed, you'll get working code for the flawed approach rather than the correct solution. Everything stays in the browser. You bring your own API key, requests go directly to your chosen provider, and there is no backend sitting in the middle. The extension records input, output, and reasoning tokens for every request and computes a Resource Unit score. RU accounts for model size and weights generated output more heavily than prompt tokens, since generation is autoregressive while prompt processing is largely parallel. The goal is not to be a perfect measure of compute, but to provide something more meaningful than raw token counts when comparing AI usage across different models. I originally built it for my own LeetCode practice, but I found three interesting use cases: Learning: AI implements your reasoning instead of silently correcting it. Iteration: Multiple approaches become runnable implementations immediately, making comparison much faster. Hiring: Resource Unit budgets could let candidates use AI within measurable limits instead of treating AI as all or nothing. I'd especially appreciate feedback on two ideas: Does constrained prompting make AI more useful for learning? Does Resource Units seem like a more meaningful metric than raw token counts? Thanks for checking it out, and I'm happy to answer any questions about the implementation or the design decisions behind it.

Pseudo

Constrained AI coding for LeetCode practice

Constrained AI coding for LeetCode practice