In this tutorial, we introduce the implementation of the Tinidev class, a minimum yet powerful AI code generation tool that uses Gemini API to transform simple application ideas in simple, structured application. Designed to run in a notebook easily, Tinidev follows a clean three-phase workflow-plan → Files → Code-Code to ensure compatibility, efficiency and modular design. Creating a web interface, python backend or utility script, Tinidev allows users to describe their project in natural language and receive ready-to-run codes files, automatically preserved to the generate and organized directory. This creates an ideal starting point to learn how it can help in fast prototyping or AI development works.
import google.generativeai as genai
import os
import json
import re
from pathlib import Path
from typing import List, Dict
We start by importing libraries needed for the Tinidev Code Generator. Google.Generati is used to communicate with Gemini API, while standard libraries such as OS, JSOn and RE support file handling and text processing. Typing paths and types of signals ensure clean file performance and better code reading.
class TinyDev:
"""
TinyDev: A lightweight AI code generator inspired by smol-dev
Uses Gemini API to generate complete applications from simple prompts
Follows the proven three-phase workflow: Plan → Files → Code
"""
def __init__(self, api_key: str, model: str = "gemini-1.5-flash"):
genai.configure(api_key=api_key)
self.model = genai.GenerativeModel(model)
self.generation_config = {
'temperature': 0.1,
'top_p': 0.8,
'max_output_tokens': 8192,
}
def plan(self, prompt: str) -> str:
"""
Phase 1: Generate project plan and shared dependencies
Creates the foundation for consistent code generation
"""
planning_prompt = f"""As an AI developer, you’re building a tool that automatically generates code tailored to the user’s needs.
the program you are writing is based on the following description:
{prompt}
the files we write will be generated by a python script. the goal is for us to all work together to write a program that will write the code for the user.
since we are working together, we need to understand what our shared dependencies are. this includes:
- import statements we all need to use
- variable names that are shared between files
- functions that are called from one file to another
- any other shared state
this is the most critical part of the process, if we don't get this right, the generated code will not work properly.
please output a markdown file called shared_dependencies.md that lists all of the shared dependencies.
the dependencies should be organized as:
1. shared variables (globals, constants)
2. shared functions (function signatures)
3. shared classes (class names and key methods)
4. shared imports (modules to import)
5. shared DOM element ids (if web project)
6. shared file paths/names
be EXHAUSTIVE in your analysis. every file must be able to import or reference these shared items."""
response = self.model.generate_content(
planning_prompt,
generation_config=self.generation_config
)
return response.text
def specify_file_paths(self, prompt: str, shared_deps: str) -> List(str):
"""
Phase 2: Determine what files need to be created
"""
files_prompt = f"""As an AI developer, you’re building a tool that automatically generates code tailored to the user’s needs.
the program:
{prompt}
the shared dependencies:
{shared_deps}
Based on the program description and shared dependencies, return a JSON array of the filenames that should be written.
Only return the JSON array, nothing else. The JSON should be an array of strings representing file paths.
For example, for a simple web app you might return:
("index.html", "styles.css", "script.js")
For a Python project you might return:
("main.py", "utils.py", "config.py", "requirements.txt")
JSON array:"""
response = self.model.generate_content(
files_prompt,
generation_config=self.generation_config
)
try:
json_match = re.search(r'\(.*?\)', response.text, re.DOTALL)
if json_match:
files = json.loads(json_match.group())
return (f for f in files if isinstance(f, str))
else:
lines = (line.strip() for line in response.text.split('\n') if line.strip())
files = ()
for line in lines:
if '.' in line and not line.startswith('#'):
file = re.sub(r'(^\w\-_./)', '', line)
if file:
files.append(file)
return files(:10)
except Exception as e:
print(f"Error parsing files: {e}")
return ("main.py", "README.md")
def generate_code_sync(self, prompt: str, shared_deps: str, filename: str) -> str:
"""
Phase 3: Generate code for individual files
"""
code_prompt = f"""As an AI developer, you’re building a tool that automatically generates code tailored to the user’s needs..
the program:
{prompt}
the shared dependencies:
{shared_deps}
Please write the file {filename}.
Remember that your job is to write the code for {filename} ONLY. Do not write any other files.
the code should be fully functional. meaning:
- all imports should be correct
- all variable references should be correct
- all function calls should be correct
- the code should be syntactically correct
- the code should be logically correct
Make sure to implement every part of the functionality described in the program description.
DO NOT include ``` code fences in your response. Return only the raw code.
Here is the code for {filename}:"""
response = self.model.generate_content(
code_prompt,
generation_config=self.generation_config
)
code = response.text
code = re.sub(r'^```(\w)*\n', '', code, flags=re.MULTILINE)
code = re.sub(r'\n```$', '', code, flags=re.MULTILINE)
return code.strip()
def create_app(self, prompt: str, output_dir: str = "/content/generated_app") -> Dict:
"""
Main workflow: Transform a simple prompt into a complete application
"""
print(f"🚀 TinyDev workflow starting...")
print(f"📝 Prompt: {prompt}")
print("\n📋 Step 1: Planning shared dependencies...")
shared_deps = self.plan(prompt)
print("✅ Dependencies planned")
print("\n📁 Step 2: Determining file structure...")
file_paths = self.specify_file_paths(prompt, shared_deps)
print(f"📄 Files to generate: {file_paths}")
Path(output_dir).mkdir(parents=True, exist_ok=True)
print(f"\n⚡ Step 3: Generating {len(file_paths)} files...")
results = {
'prompt': prompt,
'shared_deps': shared_deps,
'files': {},
'output_dir': output_dir
}
with open(Path(output_dir) / "shared_dependencies.md", 'w') as f:
f.write(shared_deps)
for filename in file_paths:
print(f" 🔧 Generating {filename}...")
try:
code = self.generate_code_sync(prompt, shared_deps, filename)
file_path = Path(output_dir) / filename
file_path.parent.mkdir(parents=True, exist_ok=True)
with open(file_path, 'w', encoding='utf-8') as f:
f.write(code)
results('files')(filename) = code
print(f" ✅ {filename} created ({len(code)} chars)")
except Exception as e:
print(f" ❌ Error generating {filename}: {e}")
results('files')(filename) = f"# Error: {e}"
readme = f"""# Generated by TinyDev (Gemini-Powered)
## Original Prompt
{prompt}
## Generated Files
{chr(10).join(f'- {f}' for f in file_paths)}
## About TinyDev
TinyDev is inspired by smol-ai/developer but uses free Gemini API.
It follows the proven three-phase workflow: Plan → Files → Code
## Usage
Check individual files for specific usage instructions.
Generated on: {os.popen('date').read().strip()}
"""
with open(Path(output_dir) / "README.md", 'w') as f:
f.write(readme)
print(f"\n🎉 Complete! Generated {len(results('files'))} files in {output_dir}")
return results
The Tinidev class contains the full reasoning of the AI-powered code generator using Gemini API. It applies structured three-phase workflow: first, it analyzes the user prompt to generate shared dependence; Next, it recognizes what files are needed for the application (specify_file_path); And in the end, it generates a functional code individually for each file (generate_code_cinsey). Create_App method brings everything together by orchesting the full application generation pipeline and saving the results including code files and detailed redem, offering a full, ready-made application scorpio from a single prompt.
def demo_tinydev():
"""Demo the TinyDev code generator"""
api_key = "Use Your API Key here"
if api_key == "YOUR_GEMINI_API_KEY_HERE":
print("❌ Please set your Gemini API key!")
print("Get one free at: https://makersuite.google.com/app/apikey")
return None
tiny_dev = TinyDev(api_key)
demo_prompts = (
"a simple HTML/JS/CSS tic tac toe game",
"a Python web scraper that gets the latest news from multiple sources",
"a responsive landing page for a local coffee shop with contact form",
"a Flask REST API for managing a todo list",
"a JavaScript calculator with a modern UI"
)
print("🤖 TinyDev - AI Code Generator")
print("=" * 50)
print("Inspired by smol-ai/developer, powered by Gemini API")
print(f"Available demo projects:")
for i, prompt in enumerate(demo_prompts, 1):
print(f"{i}. {prompt}")
demo_prompt = demo_prompts(0)
print(f"\n🎯 Running demo: {demo_prompt}")
try:
results = tiny_dev.create_app(demo_prompt)
print(f"\n📊 Results Summary:")
print(f" 📝 Prompt: {results('prompt')}")
print(f" 📁 Output: {results('output_dir')}")
print(f" 📄 Files: {len(results('files'))}")
print(f"\n📋 Generated Files:")
for filename in results('files').keys():
print(f" - {filename}")
if results('files'):
preview_file = list(results('files').keys())(0)
preview_code = results('files')(preview_file)
print(f"\n👁️ Preview of {preview_file}:")
print("-" * 40)
print(preview_code(:400) + "..." if len(preview_code) > 400 else preview_code)
print("-" * 40)
print(f"\n💡 This uses the same proven workflow as smol-ai/developer!")
print(f"📂 Check {results('output_dir')} for all generated files")
return results
except Exception as e:
print(f"❌ Demo failed: {e}")
return None
Demo_Tinidev () displays Tinidev’s capabilities by running a predefined demo using one of the many samples such as the TIC TAC TAC toe game, or producing Python news scrapers. It starts the Tinidev class from Gemini API Key, selects the first prompt from the list of project ideas, and guides the user through the full code generation pipeline, including the shared dependence, file structure and planning of the generating code. After execution, it summarizes the output, previews the sample file, and points to the directory where the entire application has been saved.
def interactive_tinydev():
"""Interactive version where you can try your own prompts"""
api_key = input("🔑 Enter your Gemini API key: ").strip()
if not api_key:
print("❌ API key required!")
return
tiny_dev = TinyDev(api_key)
print("\n🎮 Interactive TinyDev Mode")
print("Type your app ideas and watch them come to life!")
while True:
prompt = input("\n💭 Describe your app (or 'quit'): ").strip()
if prompt.lower() in ('quit', 'exit', 'q'):
print("👋 Goodbye!")
break
if prompt:
try:
results = tiny_dev.create_app(prompt, f"/content/app_{hash(prompt) % 10000}")
print(f"✅ Success! Check {results('output_dir')}")
except Exception as e:
print(f"❌ Error: {e}")
print("🎬 TinyDev - AI Code Generator Ready!")
print("Inspired by smol-ai/developer, powered by free Gemini API")
print("\nTo run demo: demo_tinydev()")
print("To try interactive mode: interactive_tinydev()")
Interactive_Tinidev () function allows users to generate applications from their custom prompts in real time. After entering a valid Gemini API key, users can describe any application idea, and Tinidev will automatically develop the entire project, code, structure and supporting files. The user continues in the process loop up to ‘leave’ types. This interactive mode enables experiments and rapid prototypes from natural language descriptions.
Finally, to ingring the demo_Tenidev (), runs Tinidev’s predetermined performance using a sample application prompt. It goes through the full workflow, planning, file structure creation and code generation, to show that the tool creates a perfect app with a simple idea.
In conclusion, the Tinidev class shows the possibility of using AI to automate application scaffolding with significant accuracy and efficiency. By breaking the code generation process into intuitive stages, it ensures that the output is logically sound, well structured and aligned with the objective of the user. Even if you are looking for new app ideas or want to accelerate development, Tinidev provides a lightweight and user-friendly solution powered by Gemini models. It is a practical tool for developers to integrate AI into their workflow without unnecessary complexity or overhead.
Check Note here. All credit for this research goes to researchers of this project. Also, feel free to follow us Twitter And don’t forget to join us 100 k+ ml subredit And subscribe Our newsletter.
Sana Hassan, a consulting intern at MarktecPost and IIT Madras, is enthusiastic about applying technology and AI to overcome real-world challenges. With more interest in solving practical problems, it brings a new perspective to the intersection of AI and real life solutions.
