The 10 Levels of AI-Driven Development
Superclaude = claude --dangerously-skip-permissions
Type that into a terminal and claude will go fully agentic and fully autonomous. **Spoiler Alert** REALLY know at a deep level what you doing and don't come crying to me if you don't!
The transition from being a manual coder to a system architect represents the most significant productivity jump in my forty+ years of electronics development. Putting some thought into it I feel there are ten distinct levels of AI integration as I've described below. Developers can move beyond writing lines of code to focus on high-level design and supervision.
This blog post covers my thinking of this evolution in which I begin at Level 0 with "vibe coding," where developers use AI chats for snippets, and progresses to spec-driven development (Level 1), where the human never touches code directly, only the requirements. As projects grow, developers adopt Functional Specification Documents (FSDs) in Markdown as a "single source of truth" (Level 2) and move to specialized agents (Level 3) like Claude Code or Gemini CLI that can edit individual lines across entire project folders rather than just single files. Security and history are managed through GitHub integration (Level 4), while productivity accelerates by granting agents full execution control (Level 5) and isolating them in Virtual Machines with "YOLO" permissions (Level 6), allowing them to run terminal commands and install software autonomously. Advanced levels involve managing parallel projects via internal memory files (Level 7) and CLI integration (Level 8), where agents compile, upload, and fix bugs in real-time by reading serial logs. Finally, by using automated data sheet analysis (Level 9) to generate libraries from raw PDFs, developers can move beyond the Arduino environment into professional frameworks like ESP-IDF (Level 10) because the AI handles the steep learning curve. Ultimately, this shift means a developer's primary tasks become writing specifications, supervising progress, and verifying results, allowing them to focus on high-level design and thorough testing instead of manual line-by-line coding
• Level 0: Vibe Coding This entry-level approach involves asking an AI chat for a code snippet or a complete sketch, then copying and pasting it into your IDE while "hoping for the best". While it is a common first step, it is inefficient because your "copy and paste fingers" will eventually need a holiday.
• Level 1: Spec-driven Development In this stage, you define your requirements in plain language and let the AI generate the code. The key difference here is that you never touch the code directly; if something fails, you update the specification rather than the code and let the AI regenerate the file.
• Level 2: Functional Specification Documents (FSD) To avoid messy chat windows, you move your specifications into structured Markdown (.md) files. Markdown is the "perfect" format because it is readable by humans and easily parsed by AI models, serving as a single source of truth for the project.
• Level 3: Agents Moving beyond simple chat interfaces, agents like Claude Code, Gemini CLI, or Codex can see your entire project folder. Unlike standard chats that require complete documents, these agents can edit individual lines, allowing them to handle long program files without overloading the AI's "context window" (its active memory).
• Level 4: GitHub Integration Once you use agents that can edit your files, version control via GitHub becomes essential. It protects you from accidental file corruption if the AI makes a mistake and provides a history that becomes vital as projects grow in complexity.
• Level 5: Full Execution Control At this level, you move past "sandboxed" agents that can only read files. Instead, you allow agents to run terminal commands, install dependencies, and execute scripts. This prevents the "terrible" scenario where you have to manually execute the AI's instructions like an "employee of your agent".
• Level 6: Development VMs To balance speed and safety, the agent is isolated within a Virtual Machine (VM). By granting the agent "YOLO" (You Only Live Once) or "dangerously skip permissions," it can execute entire workflows—like installing software or configuring environments—without asking for permission at every step.
• Level 7: Parallel Projects Using a VM and a terminal allows for managing multiple projects simultaneously. Agents can store and resume their internal knowledge files (like claude.md or agents.md) within the project directory, allowing them to restore their "context" and hit the ground running when you return to a project.
• Level 8: CLI Integration By using the Arduino CLI or PlatformIO CLI, the agent can compile, upload, and read serial logs in real-time. This allows the AI to see its own compilation errors or runtime bugs and automatically adjust the code until the project behaves as intended.
• Level 9: Automated Data Sheet Analysis Instead of manually reading 100-page PDF data sheets to find register values or multiplication factors, you simply place the raw PDF into the project directory. The AI can analyse the data sheet and generate a working library for a new chip in minutes.
• Level 10: Beyond Arduino The final level involves moving to professional frameworks like ESP-IDF. Because the AI handles the steep learning curve and framework complexities, the simplicity of the Arduino IDE is no longer a requirement, allowing for more complex and powerful professional development.
To me Claude from Anthropic is currently the most effective large language model (LLM) and agent for electronics development. The field changes rapidly and rankings may shift, but in my daily work, Claude outperforms both Google’s Gemini and OpenAI’s Codex. Claude makes significantly fewer errors when editing project files. I've had other models completely corrupt a main project file during the editing process. Beyond simple code generation, Claude Code is the most capable agent for full execution control. It can reliably run terminal commands, install necessary software, and manage complex workflows on Unix (MacOS), Linux and Windows.
Claude's is also able to handle highly technical troubleshooting and data analysis that previously required human expertise. For instance, I have had Claude successfully debug a third-party library by adding serial print statements and analysing logs to find a fix, a level of automation akin to "completed staff work". Additionally, Claude is able to analyse raw PDF data sheets for complex hardware, such as smart meters, allowing it to generate entire functional libraries in minutes without my having to read the documentation myself. To streamline my workflow even further...
I uses a custom command called "super code" to launch Claude with "dangerously skip permissions"
(hence the tag line at the top of this post),
(... or YOLO mode), allowing the agent to solve problems autonomously while I focus on high-level supervision.
