Up till this point, I've just been using AI for one-shot scripts or auto-complete suggestions, which it does well at. But I hadn't had a chance to really dig into AI powered development (i.e. just using prompts, no writing code, a.k.a Vibe Coding) on complex projects.
So when I found myself with a spare few hours over the weekend, I thought I would try AI assisted development to create a DOCX->HTML convertor in C++, compiled to WASM, and callable from within a web browser. The resulting HTML needed to be valid, and all CSS styles inlined for portability.
Initially, I was impressed with what it was able to come up with (using Claude Sonnet 4.5 Premium Requests via Github Copilot). It pulled in the dependencies it needed to parse, process, and convert the DOCX file (a structured set of XML files), and wired it altogether.
Within a couple of prompts, I had a working version that did a fairly ok job on some basic DOCX sample files (paragraphs of text with bold, italic, underlines, font size, etc), along with a functional example application for testing. Things were looking good! But there was much left to implement (images, charts, tables, and more).
Sadly, things started to go sideways very quickly. Among a myriad of issues I began to encounter over the following couple of hours were:
- The AI insisted on wasting time writing multiple large detailed Markdown files for every change, even though I had explicitly commanded it not to multiple times.
- I'd ask it to make a simple change, and for whatever reason, it would decide that variable names it had previously written were no longer good enough, and would change them all arbitrarily, leading to unnecessarily chunky git diffs
- It would inject code that was clearly invalid (e.g. XML tags into C++ code), and then spend time working out why compiling wasn't working. Most times it would find and fix the syntax errors, but sometimes it would get stuck, undo out the changes it had done, and start over completely.
- At times, whether because the codebase no longer all fit within the AI context window or for some other reason, it would start to duplicate things that already existed. And when told to tidy up the duplication, it would spend far too long making large sweeping changes that broke things.
- I would ask it to compare the code with the DOCX spec reference sheet I provided it, find ALL missing features, and implement them. But it always chose to ignore the word "ALL", fixing only one or two things, then requiring me to submit another prompt.
Overall, these issues led to many extra prompts/premium requests, wasted time, and higher costs. At a certain point, AI was being more of a hinderance than a help. And some things it just never got right.
So my verdict for now is that AI is great for one-shot scripts, auto-complete snippets, and short bursts (<30m) on new projects. But the moment the size of the context it needs to know grows too large, at least for now, AI doesn't seem to cope.
Top comments (0)