My AI Coding Workflow + Free APIs: How I DO AI CODING! (RooCode, Gemini 2.5, T3 Stack, Requesty)

AI Coding Workflow: From Idea to App

Key Concepts:

• AI-assisted coding workflow
• Model selection (Gemini 2.5 Pro, Deepseek V3)
• API key management (Requesty)
• R code integration
• Frontend framework (Next.js with T3 stack)
• Backend database (Superbase)
• Prompt engineering
• Token optimization
• AI coding tools (R code, Kodium)

1. Idea Brainstorming and Mockup Creation

• The process starts with an idea, such as creating an app to manage LLM benchmarks.
• Instead of directly prompting a model, a basic mockup of the desired UI is created using tools like Figma or Excalidraw.
• The mockup is then converted into a better-looking UI variant using prompts in ChatGPT (referencing the GPT40 designer video for specific prompts).

2. Model Selection and API Key Management

• Model Choice: Gemini 2.5 Pro is the primary model due to its cost-effectiveness and 1 million token context window. Deepseek V3 is used for simpler tasks to save money. Newer models are also explored.
• Problem: Managing multiple API keys and credits across different models is tedious.
• Solution: Transition from OpenRouter to Requesty for API key management.
  • Reasons for Switching to Requesty:
    • $1 sign-up credit and $5 extra credit for top-ups over $5.
    • Seamless integration with R code.
    • Improved dashboard with detailed usage statistics (model speed, language used, etc.).
    • API key-based settings features.

3. Requesty Features and Benefits

• Logging Feature: Logs input and output data for public applications, enabling monitoring and fine-tuning of smaller local models.
• Usage Per Key: Allows tracking token consumption across different applications (e.g., bolt DIY, clin, rue code, custom apps).
• Fallback Models: Sets up a chain of fallback models in the policy, ensuring requests are fulfilled even if the primary model fails (e.g., using free Gemini 2.5 Pro with a paid version as fallback).
• Custom Prompt: Adds a custom system prompt to every request made through an API key, overriding the system prompts of applications like R code.
  • GOU Coder System Prompt: A pre-built system prompt for R code that limits output to code, optimizing for less text and resulting in approximately 90% token savings.
  • MCP Prompt Removal: Options to remove MCP prompts for additional token savings (30% and 10% respectively).
• Token Savings: Requesty helps save on token costs, and the $5 free credit is a bonus.
• Cache Reliability: The model cache breaks less often in Requesty, further contributing to cost savings.

4. R Code Integration and Project Initialization

• R Code Setup: R code is configured with Requesty as the provider, selecting the desired model (e.g., Gemini 2.5 Pro free) and setting up fallback models in Requesty.
• Project Initialization:
  • Expo Mobile Apps: Basic expo starter kit.
  • Next.js Apps: T3 stack is preferred for type safety and TRPC integration. Alternatively, a basic Next.js app can be used.
• Initial Prompt: The initial prompt removes pre-built pages, keeps the main homepage, removes all elements from it, and changes the project title to the desired name (e.g., "king bench").

5. UI Replication and Development

• The generated image of the UI is given to the AI coder (R code) to replicate.
• Gemini 2.5 Pro is effective at replicating the UI.
• Further prompting and manual coding are required to complete the app.
• Recommendation: Learning the programming language (e.g., React, Flutter) is essential for ensuring code quality and preventing errors.

6. Backend and Tooling

• Database: Superbase is the primary database, with Firebase as an alternative. R code is used to configure the database by providing endpoint and key details.
• MCP Servers: The fetch tool and serper search tool are used for searching new libraries, documentation, and scraping pages.
• Autocomplete: Kodium is used for autocomplete.
• Windsurf and Cursor: Not used due to model truncation and limited configuration options.

7. Conclusion

• The AI coding workflow involves a combination of model selection, API key management, prompt engineering, and manual coding.
• Requesty is a valuable tool for managing API keys, optimizing token usage, and setting up fallback models.
• Learning the programming language is crucial for ensuring code quality and preventing errors.
• The presented workflow is cost-effective, with many tools offering free tiers.