The Ai thread

[tomO2013]

This was claude - paid account with Claude Pro sub.

Everything (and i mean everything) was written by Sonnet 4.5.

Including all the documentation; all i did was tell it what to make, and to write in cross platform C.

i'd never used github before (only local git), claude helped me connect the git repo to itself and publish to it.

the first version of the app was a 1 shot prompt. I added the feature for XML/JSON/CSV export and that was a one shot.

It was on the internet inside of 45 minutes as a cross platform app and that included me walking across teh street to go grab a coffee, sit back down at my desk and finish it.


Last year i did the same thing in powershell (without export formats) using chatGPT. Took a couple of hours of back and forth bugfixing and getting it right.

12 months later, its a 1 shot, tested, cross platform including upload to github with full user documentation.... in under an hour, including me learning to create the github repo and add it to claude...

• 2 years ago i would be looking for an existing tool, sorting through shareware paid stuff, etc.
• 1 year ago i got chatgpt to help and it took a couple of hours
• yesterday ... if i just wanted a tool to run locally and didn't care about publishing it... 15 minutes?

Claude pro costs like $20/month... its constantly helping save me time.

I run a software development group at a large organization, having come from a development and software architect background myself. We’ve been using Claude 4.5, Copilot, and agents heavily. I agree that as of today (Jan 2026), Claude Paid is the most useful for my personal tastes. It requires fewer corrections, though it still makes occasional bad decisions.

I do think that your mileage will definitely vary depending on the underlying tech stack and how well trained your particular model is. We recently shifted our internal development from C#/Node/React to Python/Django/HTMX and Alpine.js.
While Claude and CoPilot in general are handy research tools (similar to a quicker stack exchange google search), we found it struggled with configuration for HTMX-oriented projects initially. However, the community is filling that gap quickly—for example, this config adaptation regarding Django/HTMX skills:
https://forum.djangoproject.com/t/i...-forms-testing-htmx-drf-celery-and-more/43862

My takeaways on AI code generation:

1. Providing a Sandbox environement VS Production: AI is an excellent "sounding board" for mocking up simple UIs or exploring technical ideas that have been well trodden before. However, it does not replace the need for fundamental knowledge of the underlying platforms. Claude hallucinates. Nothing goes into our production environment without full human oversight and accountability.
2. The "Drunken Coding Partner" Effect: Coding with AI is like pair-programming with a drunken partner. It’s helpful, but you have to watch it like a hawk because it can get things authentically and confidently wrong.
3. Enterprise Maturity: Most models that I have used are not yet ready to handle enterprise-grade concerns like security, scalability, performance considerations and complex logging without heavy hand-holding. Anywhere you see hand-holding expect significant time investment.
   Where they do excel is with isolated tasks (like parsing a log file). They tend to struggle with asks that start to straddle into more complex territory like holistic architecture!
4. The Efficiency Myth: There is a belief among (non-tech-savvy) executives that AI will slash costs by replacing developers. This ignores the reality that software delays (at least in my experience) are rarely caused by typing speed. Rather they are often caused by unclear business requirements or scope creep. AI cannot fix a lack of understanding of the core business problem.

Ultimately, if I could wave a magic wand to make most software development projects go quicker at my own company, I truly believe that empathy will become a superpower that differentiates developers of the future. As I tell my own team, understanding the user and the problem is the one skill that keeps developers relevant, and it’s something AI cannot replicate.

Just my 0.02.

(disclaimer: the above is my own honest opinion and should not be held as being that of my employer).