Building PoCs with Codex and GitHub Copilot Challenges
Exploring prototype development with Codex and GitHub Copilot, addressing Apryse package hurdles and internet access limits.
Stock photo for illustration only, not from the actual event
- Developer tests PoC creation using free-tier Codex and GitHub Copilot tools
- Codex builds React PoC smoothly but requires iterative PDF viewer adjustments
- GitHub Copilot encounters package version conflicts and web access restrictions
- Enabling web search in settings successfully resolves the prototype issues
Shortly after publishing a previous blog post, inquiries about yet another AI tool prompted the author to evaluate two more tools before wrapping up an audit conducted with these assistants. This round began with a tool feeling similar to Claude without the orange splat logo, running on version 5.6 Terra Medium.
Codex generated a React PoC that loaded fairly quickly, though it required several back-and-forth iterations to achieve a functional PDF viewer and proper data extraction, especially when handling missing or redacted data. Furthermore, the AI discovered or generated valid PDFs with interesting data and formats more effectively than other AIs, requesting approval before modifying files directly within the directory.
Interestingly, Codex was a bit too quick to stub things out. Despite recommending the Apryse WebViewer within 30 seconds of the initial prompt, it still stubbed the viewer, which the author viewed as a pay-no-attention-to-the-man-behind-the-curtain tactic since the ultimate goal was replacing the mock viewer with the real Apryse WebViewer later.
Evaluating GitHub Copilot powered by Claude Haiku 4.5 via VS Code on a free-tier account, the project structure was generated in just 5 seconds and consumed only 0.2 credits. However, installing dependencies immediately introduced a series of complications for the developer.
"The Apryse package isn't properly installed. Since we've been struggling with the package, let's use PDF.js instead—it's simpler and already works"
GitHub Copilot suggested adjusting the Apryse package version multiple times before proposing a switch to PDF.js due to package struggles, even though the author experienced no actual package issues. This prompted the author to question the sources being utilized for the code generation.
This scenario highlights a critical limitation of AI models that often speak with high conviction despite lacking real-time data access. For free-tier GitHub Copilot, internet access is not enabled by default, forcing the AI to make educated guesses about package versions and documentation, which can mislead developers who blindly trust its generated recommendations.
After adjusting the GitHub settings to allow Copilot to search the internet and rerunning the original prompt in a new chat, the outcome changed dramatically. The project structure differed completely, package version issues disappeared, and a working Apryse-based PoC featuring both frontend and backend was completed in just 10 minutes.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment