All Projects

Classway

Written by Saif Pasha. Originally published on saifpasha.com.

Classway is an AI humanizer and answer tool for students, built for a young founder who wanted students to get a straight answer in their own voice instead of something that read like a machine wrote it. We designed and built it as a Chrome extension and a web app. Today it has more than 222,000 users who have asked it over three million questions.

Background

A young founder came to us in the early days of ChatGPT with a clear read on the problem. Students were already using AI to answer questions and write essays, and what came back read like AI, which got flagged by teachers and detection tools. He wanted a tool that gave the same help but in a voice that sounded like a student wrote it.

He was starting out, so the product had to be cheap to run as well as cheap to build. Every call to an outside AI model would be paid again on every single answer, for as long as the product lived.

The problem: AI detection

The standard approach at the time was heavy. Generate an answer, send it to a detector to score how much it reads as AI, run it through a separate humanizing service to rewrite it, then send it back to the detector to confirm it passed. That is four paid calls for one student question.

We tested that approach before building anything. AI detectors were unreliable in those days. The same paragraph could score fully human on one detector and fully AI on the next. So every version of the prompt was checked against five detectors at once. The best humanizing service on the market turned out to be nothing special under the hood. It was just a language model with a well-written prompt of its own.

What we built

That finding changed the whole design. If a humanizing service was only a model with the right instructions, Classway did not need four calls or a separate detector. We rebuilt the pipeline four times, each version with fewer calls than the last, until it came down to one call that writes the answer and humanizes it at the same time.

The prompt took longer to get right than anything else. Saif rewrote it against all five detectors until a pattern held: polished, evenly punctuated writing is what gives a model away, and a student's real writing does not look like that. The final prompt asks for a grade level, tested from 8th grade through 12th, and for the kind of small slips a student actually makes, a missing comma, an imperfect sentence, a point repeated the way a student circles back to it. With that built in, the rewrites scored above 95 percent human on every detector tested. At that point the detector calls came out of the product entirely. Every answer started as AI text, so there was no reason to pay a detector to confirm it, and the score a student saw was set by what the humanizer had already proven in testing, not a live check on their answer.

How it works

One call instead of four

The old design made four paid calls per question. What shipped is one call to OpenAI's chat completion models, whether Classway is answering a question or rewriting text a student pastes in. The server holds the API key, picks the model, and streams the response straight back to the extension or the website.

Testing against five detectors

Every change to the prompt had to hold up across five different AI detectors at once, since no single detector could be trusted on its own. A change only shipped once it passed all five. That testing is where the humanizer's accuracy came from, not any one tool.

Writing like a student, not a professor

The prompt names a grade level and asks for the kind of imperfect writing a real student produces: grammar that slips occasionally, commas left out where they belong, a point mentioned again the way someone circles back mid-thought. Those small details moved the detector scores more than anything else tried.

Keeping up with new models

Each new OpenAI model release changed how the underlying model wrote, which meant a prompt tuned for one model would drift on the next. Saif retuned the prompt after every release, from GPT-3.5 Turbo through GPT-4 and GPT-4 Turbo to the model Classway runs on today, so the score and the cost held steady across every upgrade.

The result

Students moved almost entirely to the web app, so the team dropped the Chrome extension rather than maintain a version few people used. Classway has more than 222,000 users today and has answered over three million questions since launch.

You have outdone yourself Saif

Ali Debba, founder of Classway and NiggleAI

The same founder came back months later with a second product. We built that one too, NiggleAI, a study tool that turns documents and lectures into notes, flashcards and quizzes.

Who did what

Burhan Pasha was the client lead on this project and owned delivery from the first call through launch. Saif Pasha designed Classway's brand and every screen, built the humanizer, and tuned the prompt through every model change. The Silverthread Labs team built the extension and the web app around it.

If you're building an AI product that has to hold up once real users arrive, talk to Burhan.

// NEXT STEP

Building something like this?

Tell me what you're trying to build. The first call ends with what to build first, whether or not you hire us.