↓Skip to main content

Draw Me a Story: AI Creates & Reads Tiny Tales from Five Simple Prompts

·4 mins

I built a tiny creative assistant that asks for five things — a character, color, place, feeling, and an action — then crafts a short, age-appropriate story, reads it aloud, and generates a playful image to finish.

This project is a small experiment in constrained creative prompting, safety filtering for young audiences, and combining LLMs with TTS and image generation for an end-to-end, delight-first experience.

✨ Demo #

✨ How it works (happy summary) #

Give the bot five inputs:

  • Character: e.g., “a tiny dragon”
  • Color: e.g., “yellow”
  • Place: e.g., “a cozy treehouse”
  • Feeling: e.g., “curious”
  • Action: e.g., “bakes a cake”

The chatbot then:

  1. Builds a short 2–3 minute story with simple vocabulary and gentle pacing for toddlers.
  2. Applies a safety + content filter to ensure the language, themes, and imagery are age-appropriate.
  3. Sends the cleaned final text to a TTS system and produces an audio file (MP3/OGG).
  4. Creates a bright, cartoon-style image based on the same inputs to display at the end.

🧩 Example #

Inputs:

  • Character: a tiny dragon
  • Color: yellow
  • Place: a cozy treehouse
  • Feeling: curious
  • Action: bakes a cake

Generated story (trimmed):

Little Sunny was a tiny yellow dragon who lived in a cozy treehouse high in the branches. Sunny was very curious — every morning they peeped out the window to see what new giggles the day would bring. One sunny morning, Sunny decided to bake a cake for a new friend. They mixed flour, a little pretend sugar, and a sprinkle of laughter. “Whoosh!” said Sunny as the oven made a friendly humming sound. When the cake was ready, Sunny shared a slice and a big, warm dragon hug. The End.

The same text is passed to the TTS engine and a cheerful, slightly slower reading is produced to suit toddlers.

🔧 Implementation notes (technical details) #

This section mirrors the technical approach from a similar project I wrote about in “From Web to Wonder” — but simplified for this five-prompt assistant.

  • Architecture & UI: A lightweight server orchestrates story generation, TTS, and image creation. For demos and input collection I built a small Gradio UI that collects the five prompts, previews the generated story text, plays the TTS audio, and displays the final image.
  • LLM model API: The core story generation uses an LLM model API (hosted or cloud provider). I call the model via its HTTP SDK, requesting a structured JSON response (e.g., {"title":"...","story":"...","safetyScore":0.0}). Using the API lets the app validate responses, retry with adjusted prompts, and log examples for analysis.
  • AI tools: Separate cloud/hosted tools handle TTS and image generation (or local models). I pass the sanitized story to a TTS service (producing MP3/OGG) and send a controlled image prompt to an image model that returns a cartoon-style illustration.
  • Prompting / Constraints: Prompts enforce short sentences, a small vocabulary list appropriate for toddlers, and an explicit ban on frightening or adult content. Example instruction snippet:
You are a storyteller for toddlers. Use simple words, short sentences, gentle tone. Return JSON with fields: title, story, safetyScore (0-1).
  • Safety filtering: Two-stage filtering: model-side constraints plus app-side checks against a denylist and a readability simplifier that shortens sentences and replaces rare words.
  • Model tuning: Low temperature (0.1–0.3) for consistent, child-friendly outputs; token limits keep stories brief.
  • API flow / functions: The app calls POST /generate_story (LLM API) -> validate -> POST /synthesize (TTS) -> POST /generate_image (image API). The Gradio front-end calls the server, streams the audio preview, and shows the generated image.

🛡️ Safety & design decisions #

  • All generated text goes through a two-step sanitization before being used for TTS or image prompts.
  • Prompts explicitly instruct the model to avoid real-world instructions (no cooking instructions with hot ovens, no small-object-choking risks), and to prefer playful, imaginative descriptions.
  • Images use a cartoon style and avoid photorealism for younger audiences.

📂 Code #

Implementation details, prompt templates, and the small orchestration code are available in the repository: LLM Engineering Github source code