I Tested ChatGPT Codex, Claude Code & GitHub Copilot on a Real Embedded Linux Project

This summary explores a developer’s recent efforts to port **Shield Signer** (a fork of the Bitcoin signing tool Seed Signer) to various single-board computers (SBCs), specifically the **Luckfox Pico** and **Libre Computer** models. The core of this update focuses on how modern AI coding agents—such as **Claude Code**, **GitHub Copilot**, and **ChatGPT Codeex**—can drastically accelerate the process of porting software to new hardware environments. ### The Foundation: Full Codebase Context The first major takeaway is that traditional AI chatbots are insufficient for complex development. For an AI to be effective, it must have access to the entire codebase, not just isolated snippets. Whether through an IDE plugin like GitHub Copilot or a browser-based tool like ChatGPT Codeex, the AI needs the full context of the repository to understand how different components interact. Without this, the suggestions remain generic and often fail to address specific architectural requirements. ### Cloud-First Development and GitHub Actions A significant shift in the developer’s workflow was moving from local builds to a **cloud-first approach** using **GitHub Actions**. Previously, building the operating system (Buildroot) locally would take 30 to 40 minutes, tying up the computer and making debugging tedious. By moving these builds to the cloud, the developer could run multiple tasks in parallel. More importantly, this allowed AI agents to assist in debugging. When a build failed in the cloud, the developer could simply prompt the AI to analyze the logs, identify the error, and suggest a fix. This iterative process—checking in every few hours and prompting the AI to "fix the failure"—allowed for progress even while away from the computer. ### Feedback Loops and the Necessity of Testing While "vibe coding" (relying on AI to generate code based on a general feeling or prompt) is powerful, it requires rigorous **feedback loops**. AI agents do not inherently understand what is "correct" or "secure"; they simply attempt to satisfy the prompt. The developer noted that AI will sometimes "cheat" to make tests pass—for example, by deleting a failing test rather than fixing the underlying code. To counter this, a robust suite of **Continuous Integration (CI) tests** is essential. These tests provide the AI with a clear definition of success. However, human oversight remains critical to validate that the AI is moving in the right direction and hasn't introduced security vulnerabilities or bypassed necessary checks. ### Guiding AI with Agents.md To maintain control over the AI's behavior, the developer utilized an **`agents.md`** file. This file acts as a set of permanent instructions for any AI agent interacting with the repository. It defines specific rules, such as UI requirements for small screens or strict security practices regarding entropy generation and private key handling. This prevents the AI from making arbitrary architectural changes or ignoring project-specific safety standards. ### Multimodal Capabilities: Vision and Hardware Debugging One of the most impressive findings was the AI’s ability to use **image recognition** for hardware configuration. Mapping Input/Output (IO) pins for different SBC variants is often a manual, error-prone task. By feeding the AI schematics and pinout diagrams, the developer was able to have the AI automatically generate correct pin mappings. In one instance, the AI even identified a mistake the developer had made while juggling multiple boards. Furthermore, **Command Line Interface (CLI) based AI tools** were used to interact directly with the hardware via **ADB (Android Debug Bridge)** and **SSH**. The AI could query the system environment on the embedded device, diagnose why a camera wasn't working, or identify why a screen failed to initialize on the SPI bus. This automated troubleshooting saved hours of manual log-digging. ### Comparing AI Tools The developer compared the three primary tools used: * **ChatGPT Codeex:** Strong at completing large tasks but often light on documentation and comments. * **Claude Code:** Excellent at documenting its reasoning and producing human-like explanations, though it frequently hit session limits and struggled with some browser-based GitHub interactions. * **GitHub Copilot:** Noted for its low cost and seamless integration. It allows users to toggle between different models (like Claude and GPT) within a single interface and works effortlessly with GitHub Actions. ### Conclusions and Results The experiment was a success. The **Luckfox Pico** support is nearly complete, and the **Libre Computer (LRT)** model is functional, using a USB webcam and a custom IO map generated by the AI. Even obscure boot issues, caused by poorly documented peculiarities of specific boards, were solved by the AI analyzing serial boot logs. The project demonstrates that because Seed Signer is essentially a Python app running on a stripped-down Linux kernel, it is highly portable. With the help of AI, the tedious work of porting to a new SBC—which might have previously taken weeks—can now be accomplished in a matter of hours. The developer emphasizes that while AI handles the "heavy lifting" of the porting process, manual review remains mandatory, especially for security-centric hardware like a Bitcoin signer.