Chrome Extension · MV3 React + Vite FastAPI Fine-tuned LLM

PocketZot 🐜

A desktop-mascot Chrome extension that lives on AI chat sites, coaches you toward learning-oriented prompts, and rewards good habits with ants you spend on hats. This page is a live, no-install demo of the mascot & prompt coach.

How PocketZot is trained

PocketZot judges a prompt by where it lands on Bloom's Taxonomy, the classic hierarchy of thinking skills. Asking an AI to do your lower-order thinking (recall, summarize) skews toward offloading; keeping the higher-order thinking yourself (analyze, evaluate, create) is learning-oriented. The model is fine-tuned to spot the difference.

Create "Help me design my own approach to..." learning
Evaluate "Critique my reasoning and where it breaks down" learning
Analyze "Ask me guiding questions to solve this" learning
Apply "Walk me through applying this to a new case" mixed
Understand "Summarize this chapter for me" offload-prone
Remember "Just give me the answer to..." offload-prone
  1. Label against Bloom's levels. A hand-labeled set of real prompts is tagged with the cognitive level it asks the AI to perform.
  2. Collapse to a learning signal. Lower-order levels map toward offloading, higher-order levels toward learning, giving each prompt a target label.
  3. Fine-tune GPT-4o mini. The model learns to predict the level from the prompt alone, in a single fast call, instead of relying on brittle keyword rules.
  4. Serve in real time. A FastAPI endpoint runs the classifier so the extension can score each prompt as you type it on ChatGPT, Claude, Gemini, and Perplexity.
  5. The anteater reacts. Higher-order prompts earn ants and cheers; lower-order ones get a gentle nudge to reframe, so the habit loop reinforces real learning.

The anteater is the real extension engine running unchanged. Classification in this demo is a local offline stand-in for the fine-tuned model used in the full app.