← Fred Luz
Live

Kaoshi

AI worksheet generator for teachers.

The breakthrough wasn't better AI—it was fixing the iteration loop.

I

The Problem

Teachers spend hours creating worksheets. Not the fun part—planning lessons, engaging students—but the mechanical work of generating questions, formatting, adjusting difficulty.

When ChatGPT launched, teachers flocked to it. Finally, a way to automate the drudgery. But something interesting happened: they kept using it, but they kept complaining.

A typical ChatGPT session

Create 10 math questions about quadratic equations for 10th graders
1. Solve x² + 5x + 6 = 0
2. Factor x² - 9...

"Too easy. Wrong difficulty distribution. Can't edit inline. Have to copy-paste everything and start over."

— Teacher's internal monologue

"ChatGPT gives me garbage, but at least I can edit it. The other AI worksheet tools give me garbage I can't fix."

This was the pattern: generate, review, copy-paste into Word, edit, format, repeat. The AI saved time on generation but created friction everywhere else.

II

The Constraint

Existing AI worksheet tools made the same mistake: they treated generation as a one-shot process.

But teaching is inherently iterative. A worksheet isn't "done" when the AI finishes generating—it's done when the teacher is satisfied.

The Bottleneck

Not generation speed. Iteration speed.

ChatGPT

  • Can edit output
  • Conversational iteration
  • Loses formatting
  • Copy-paste workflow

AI Worksheet Tools

  • Purpose-built interface
  • Keeps formatting
  • Can't edit the plan
  • One-shot generation

ChatGPT won not because it generated better content, but because it let teachers iterate conversationally. The cost was context loss—each prompt started fresh, formatting broke, mental models fragmented.

III

The Insight

What if we separated planning from composition?

Teachers don't want to write questions from scratch—that's what AI is for. But they do want to control the pedagogical structure.

The plan is where teacher expertise lives. Question types, difficulty distribution, cognitive levels.

Let teachers own the plan.

Let AI handle the composition.

The Two-Phase Pipeline

Planning Phase

AI generates a structured plan. Teacher reviews and edits.

Q1: Multiple choice, Easy, Knowledge
Q2: Short answer, Medium, Application
Q3: Essay, Hard, Analysis

Composition Phase

AI generates each question. Teacher watches, pauses, edits.

Generating Q1... ✓
Generating Q2... ✓
Generating Q3... ...
IV

The Craft

Building the plan editor was the hardest part. It needed to feel as fluid as editing a document, but with structured data underneath.

Every interaction was optimized for speed of iteration.

Undo/Redo (50 states)

Teachers experiment freely knowing they can revert. State history persists across sessions.

AI Suggestions

The plan critiques itself: "Add more application-level questions" or "Balance difficulty better."

Drag & Drop

Questions flow better when sequenced intentionally. Reorder by difficulty, topic, or pedagogical arc.

Pause & Resume

Generation streams in real-time. Pause mid-stream to review, then resume or regenerate.

The "Make It Harder" Problem

When teachers ask ChatGPT for a harder question, they often get the same question with more complex vocabulary—not genuinely higher cognitive demand.

Original:

"What is the main idea?"

↓ "Make this harder" ↓

"What is the central thematic conceit?"

Same cognitive level. Just bigger words.

Kaoshi's Fix

Explicit difficulty parameters the AI can't fake:

  • Bloom's taxonomy level
  • Reasoning steps required
  • Cognitive complexity factors
Kaoshi method selection

Five entry points, one consistent experience

Technical note: The LLM integration isn't special—Vercel AI SDK. The state management is where the work went. Client-side React hooks, localStorage persistence, optimistic UI. Reduced LLM abstraction from ~1,500 lines to ~200. But that's not the story.

V

The Lesson

Teachers don't want AI to replace their judgment. They want AI to accelerate their workflow while preserving their agency.

The best AI tools aren't the ones with the most advanced models. They're the ones that understand the human workflow and remove friction from the iteration loop.

Kaoshi's breakthrough wasn't Kimi-K2's 200K context window or Claude Haiku's speed. It was the realization that planning and composition are different cognitive modes—and building an interface that respects that difference.

Control beats automation.

Kaoshi generation flow

Questions appear one by one, editable at any time

Built with

Bun React TypeScript Hono Supabase Vercel AI SDK Tailwind