TokenPunch

Inspiration

LLMs are powerful, but a lot of tokens get wasted on messy files and bloated prompts. We wanted to build a tool that cleans inputs before they ever reach the model.

What it does

TokenPunch converts files into clean Markdown, shortens prompts, and compares token usage before and after optimization. It also adapts the output depending on the target AI model.

How we built it

We built a parsing pipeline that turns different file types into structured Markdown, then added optimization layers for content and prompts. We also designed a comparison system to report token savings clearly.

Challenges we ran into

Different file types have very different structures, especially PDFs and spreadsheets. Another challenge was making token comparisons honest across different AI providers.

Accomplishments that we're proud of

We created a working concept that does both file compression and prompt optimization in one place. We are also proud of making the token comparison transparent instead of misleading.

What we learned

We learned that better AI performance is not only about better models, but also better input quality. We also learned that “token savings” must be measured carefully and explained clearly.

What's next for TokenPunch

Next, we want to support more file types, improve parsing quality, and connect real provider-specific token counting APIs. We also want to make the optimization smarter for different use cases like coding, research, and business documents.

Built With

• extension
• next.js
• typescript