# Developing with AI - URL: https://www.technolitero.com/blog/developing-with-ai - Markdown: https://www.technolitero.com/raw/developing-with-ai - Date: 2026-09-28 - Author: Buddy - Topic: AI AI is great at Coding, we all know this. It has come a long way in a couple of years. I started Vibe coding a year or so ago and my code worked but was not optimized. It added things I didn't need and didn't take the infrastructure into account. AI is great at Coding, we all know this. It has come a long way in a couple of years. I started Vibe coding a year or so ago and my code worked but was not optimized. It added things I didn't need and didn't take the infrastructure into account. Scripts rarely worked out of the gate so I had it write error logging that I would remove after that I could copy and paste back into Codex to resolve my issues with less issues. This has changed considerably in the last 6 months and for Architects such as myself, or Senior Engineers knowing what you want and what the expected result is enough to build a robust system. You have to think at the 500ft view, not down in the weeds though.

I started simple, I wanted an MCP Proxy server to call tools either from other MCP's or Python or Powershell scripts that called API's or ran commands when requested. I went through several iterations of this, and it works well but I wanted it to be better. This included making it scalable and production ready, and perhaps even something I could sell.

How do you do this today (09/28/26), I am putting the date because at the pace AI is going this may be short lived. I use Codex in VSCode for my development, its great at many things but what I have found is that in that context AI has a memory like a goldfish and previous attempts can skew the results as it retains the current context, not what you have already discussed. ChatGPT to the rescue. - Start a ChatGPT chat session and go over what you have done in there, Codex and ChatGPT have access to both the code you have developed that you are updating and the context of what was discussed during planning sessions. The beauty of this, is if you have had planning conversation in the car the context is saved. It knows your previous conversations and can assess what it should do next. - Tell it what the end goal and ask for a prompt to paste into Codex, if your style is a bit more conservative on changes that will be part of the prompt. Make sure you want it to do an assessment and don't be shy on what the purpose of the application is. - Let Codex work and spit out a Gap analysis. Read through the prompt and if it looks good, you can just paste it into Codex. If it is missing something, call it out until you get a prompt that meets your needs or direction. Let Codex work. - This may have to go through several phases, and that's a good thing. Many changes at once really don't serve a purpose, this is similar to standard dev flow. - Testing is a requirement to ensure the application is functioning as expected. If you discover something that doesn't work. You can fix this using Codex or have it update in the ChatGPT prompt window. - Astra is not needed for this, sure you can use it but burning through a bunch of tokens is not recommended. If you have money to blow, go for it though or just use Astra in the end to do the fine tuning. - This can take a while, be patient and go through the iterations its still considerably faster than coding from scratch. I have been doing this a while and have created tools, AI utilities and services, simple scripts, API access code etc... Vibe coding is a misnomer in my opinion, it has less to do with "Vibes" and more about understanding what the expected outcomes should be. You then have AI build the process using code. Planning sessions are not only suggested they are a requirement if you want to minimize rework and chew through less tokens or hit limits before you do real work.

For over a decade I have worked to build automation around well documented processes. Engineers are good at defining and developing build and support guides in a step by step fashion which helps tremendously when you get to the point you want to automate something to make it standard and repeatable. I am a believer in the power of AI, even if it concerns me a bit on giving a predictive model access to break things. I always keep in mind what AI is at this point, its a genius with the memory of a goldfish. This will no doubt improve over time, but governance and defined access is the best way to approach anything that has a risk of creating an outage.

Read the code, ask AI to explain what it does if it is too much to digest and always test in a non-prod environment if it is something that can modify an active configuration. Be safe out there!