Boss Fight: CLI Chat

The Scenario

You’re building an interactive CLI chat — a terminal program that works like a mini ChatGPT. The chat combines all six building blocks from Level 1: SDK setup, model selection, system prompt, streaming and structured output.

Your chat should feel like this:

> Hallo! Wer bist Du?
Ich bin Dein AI-Assistent. Ich kann Dir bei Fragen helfen...

> /json Was ist TypeScript?
{
  "topic": "TypeScript",
  "summary": "TypeScript ist eine typisierte Erweiterung von JavaScript...",
  "keyPoints": ["Statische Typen", "Bessere IDE-Unterstuetzung", ...],
  "difficulty": "beginner"
}

> Tokens bisher: 342

Expected duration: 30-45 minutes. Create a file chat.ts and run it with npx tsx chat.ts.

This project connects all six building blocks:

User Input flows into selectModel to Model, together with System Prompt into streamText/Output.object to CLI Output

Requirements

Interactive input — The user can enter messages repeatedly (readline or process.stdin). The program runs until the user types “exit”.
System prompt (Challenge 1.6) — The chat has a defined role with rules and style. The system prompt is passed with every call.
streamText for normal responses (Challenge 1.4) — Normal messages are answered with streamText. Text appears token by token in the terminal.
Output.object + Zod schema on “/json” (Challenge 1.5) — When the user starts a message with /json, the response is generated as a structured JSON object (generateText + Output.object).
selectModel based on message length (Challenge 1.2) — Short messages (under 50 characters) use a cheap flash model. Long messages use a pro model.
Usage tracking (Challenge 1.3) — After each response, the consumed tokens are displayed. A total counter tracks the session.

Starter Code

import * as readline from 'readline';
// TODO: Import AI SDK functions (generateText, streamText, Output)
// TODO: Import provider (anthropic from '@ai-sdk/anthropic')
// TODO: Import zod (z from 'zod')
// Tip: If you only have one provider, use different models
// from the same provider -- e.g. anthropic('claude-haiku-3-5-20241022')
// for short and anthropic('claude-sonnet-4-5-20250514') for long messages.

// TODO: Definiere einen System Prompt mit Rolle und Regeln

// TODO: Definiere ein Zod Schema fuer den /json Modus

// TODO: Implementiere selectModel(message: string)

let totalTokens = 0;

const rl = readline.createInterface({
  input: process.stdin,
  output: process.stdout,
});

function askQuestion(): void {
  rl.question('\n> ', async (input) => {
    const message = input.trim();

    if (message === 'exit') {
      console.log(`\nSession beendet. Tokens gesamt: ${totalTokens}`);
      rl.close();
      return;
    }

    if (message === '') {
      askQuestion();
      return;
    }

    // TODO: Pruefe ob die Nachricht mit /json beginnt
    // TODO: Waehle das Modell basierend auf der Nachrichtenlaenge
    // TODO: Bei /json → generateText + Output.object
    // TODO: Sonst → streamText + textStream
    // TODO: Aktualisiere totalTokens

    askQuestion();
  });
}

console.log('CLI Chat gestartet. Tippe "exit" zum Beenden.');
console.log('Starte eine Nachricht mit /json fuer strukturierte Ausgabe.\n');
askQuestion();

Evaluation Criteria

Your Boss Fight is passed when:

The chat runs interactively in the terminal and reacts to user input
A system prompt defines the assistant’s role and behavior
Normal messages are output with streamText token by token
Messages starting with /json produce structured output with Output.object and Zod schema
selectModel chooses the model based on message length
After each response, the token usage is displayed
A total counter tracks the tokens for the entire session
“exit” ends the program cleanly with the total counter

Hints

Hint 1: readline and async/await

The readline callback is not directly async. But you can mark the callback as async — this works because readline ignores the callback’s return value. Alternatively you can create a separate async function handleMessage(message: string) and call it from the callback.

Hint 2: Writing stream output to the terminal

For streamText use process.stdout.write(chunk) instead of console.log(chunk), so that the text appears word by word without line breaks. After the stream, use a console.log() for the final line break. You can get the token usage from fullStream (event type finish) or via result.usage (a promise you can await after the stream).

Hint 3: /json detection and prompt extraction

Check with message.startsWith('/json') whether JSON mode should be activated. The actual prompt is then message.slice(5).trim() — everything after “/json”. For JSON mode use generateText (not streamText), because Output.object needs a complete object, not a stream.