AI-Powered Coding: A Developer's Reference Playbook

Why I Use AI for Coding

As a software developer, I use AI not only for speed but also to enhance code quality and facilitate continuous learning. This guide is a breakdown of how I integrate AI tools during real-world programming sessions.

Whether I'm exploring a new idea, writing docs, or optimizing a script, AI is part of the flow.

Coding Modes

1. Vibe Coding (Prompt-Only Code Generation)

Let the AI do all the typing. I write detailed prompts, review the output, and give feedback.

Ideal for:

• Rapid prototyping
• Spiking unfamiliar ideas
• Exploratory code
• Good vibes only

Examples:

• "Write a Python script named save_res that hits this API and saves the response as JSON in the /data directory."
• "Write another script that loads the latest file in /data and runs an analysis summary using the existing analyze.py script."

💡 Tip: It's all about context!

For coding complete applications, I use a AI to help me create a Product Requirement Document (PRD), which I then add as a markdown file in an empty project. This marks the start of the LLM's coding process for the entire application.

Oftentimes, a PRD is only the beginning.

Having multiple markdown files even before you begin vibe-coding is the best way to ensure you get great results.

For example, files like api-spec.md, accessability-guidelines.md, and user-flows.md are an excellent start before any coding actually begins.

2. Assisted Coding (Manual + AI Combined)

I write, then let AI help me refactor, document, test, and reason through problems. Remember, humans read code too!

Ideal for:

• Production-quality code
• Working with large codebases
• Learning by doing

Typical Uses:

• Suggesting better variable names (always the hardest!)
• Writing inline comments
• Refactoring for readability
• Drafting tests for new functions

Tools I Use:

• GitHub Copilot (in IDE)
• ChatGPT/Claude, Gemini, etc
• Local LLMs (great for one-off questions)

My AI Coding Stack

1. Copilot/Cline/Roo/LM Studio (Chat + Agent)

• Use for: Inline suggestions, refactoring, tests/docs
• Chat: Ask for explanations, improvements.
• Agent: Generate scripts that process or manipulate files.

2. Web LLMs for Research & Writing

• Use for: Prompt engineering, product docs, deep research
• Tools: Claude, Gemini, ChatGPT

Example workflow: Use Claude to draft the requirement docs. Use Gemini to collect background information about the product or market. Then have ChatGPT to refine and experiment with prompt syntax.

Using different models for tasks is a great way to diversify. For science, you can even give them all the same task (like writing PRDs!) and choose your favorite, or have another LLM evaluate them all and combine them into one document.

3. Vibe Coding Agents

• Tools: Replit, OpenAI Codex, Roo, Cline, Claude

Example Workflow: Generate base code using Replit. Refine against it using Codex or Claude. Analyze and extend code via Roo or Cline.

4. Local LLMs

• Use for: Utility scripts, privacy-sensitive tasks, fast feedback

Example workflow: Ask for help with error messages, syntax, or writing isolated functions. Run planning logic locally before sending to cloud-based models to save on cost.

💡 Tip

For heavy prompts (such as vibe coding, architecture, or system planning), I don't waste tokens or time on smaller models. I start with a reasoning model to establish a solid foundation, then let lighter tools refine or implement it.

Sample Use Case Workflow (Multi-Agent Style)

Goal: Build a script that makes an API call and analyzes the result.

1. Prompt GPT/Claude:
2. "Write a Python script that makes a GET request to the given endpoint and stores the response as a JSON file in ./data."
3. Agent Behavior (Cline/Roo):
4. Writes and saves the file automatically.
5. Follow-up Prompt:
6. "Now write a second script that reads the latest file in ./data, extracts metrics, and prints a basic summary."
7. Refactor + Docs (Copilot/Claude):
8. Ask for documentation and inline comments.
9. Tests (Codex or ChatGPT):
10. Prompt: "Write extensive tests for both scripts."

Final Thoughts

AI is not just a shortcut; it's a collaborator. By switching between vibe coding and assisted coding, I can go from idea to working prototype, refine it, and even explain it to someone new (human or not!) within a single coding session.