Are you using AI tools for coding and software dev?

No. I have many moral objections to the big AI companies, how they train their models and their purpose. But also, personally, I like the challenge of solving coding problems, and so I don't want to outsource the things I enjoy doing. I also don't want to depend on services that are currently heavily subsidised, but soon won't be.

So i have done some development on Apps and other tools in a couple of ways. And what you say about Claude is very much a reason I have moved down a specific path.

I have done development with Codex which out the box is simple enough. However biggest issue is it needs plugins and extras to help it work better. It is super good at using all your usage and spaffing it up the wall. Also overcomplicating the development too.

I have a specific type of build process that i use that is tried and tested and works. And for some reason newer models seem to like to ignroe my process and think they know best. For instance ChatGPT has a new model Astra 6 which is fantastic however for my build process its horrible in terms of efficient usage. It decides to take my process and think i know better i will reinvent the wheel to create this. So instead the GPT 5,5 model has been better for my purposes and a hell of alot cheaper.

From all the different tools and developments ive used Ive now rested on Hermes Agent. Started with Openclaw, tested paperclip and god knows how many others.

Its really easy to develop on those out the box just like Codex. But i have found there is far more that can be added in plugins, tools and skills with Hermes. Now with my setup I have a suite of experts to be able to query to really get things refined. skills for proper devleopment structure and able to add so many tools from github that help to map out processes etc.

Its still not perfect but its enabled me to develop so much that has been useful to me and also build new products. Some of those are discord bots for specific communities ive needed them for. Discord integration so i can use my agents from discord without any extra needs. Complete suite of CRM, ERP system that has the AI Agent integrated to chat to it to automate processes in it. Website updates. Even a fully hyper personalise cold out reach process for email and linkedin. And with the additional tools I can export them easily to have them deployed for a client.

Ive even developed with creating an App for a game (since the game doesnt give the quality of life tools) and its enabled me to develop it on my laptop using computer use and much more while using the remote gateway i have hosted on my server meaning i can have a large group of credentials it uses to work on rather than spanking all the usage on one set of credentials.

When it comes to projects a proper plan is needed and I have forced my Agent to setup planning documents it uses as canon and tracks that in the file so it doesnt lose its position. And there are a few different memory requirements i would state to have. It needs context memory (something it can search and summarises often to keep token usage low but that the AI can search for and find when needing context), regular memory so it understands its processes and how it should work and then the "Second Brain Vault" a vault of knowledge that it adds to over time to keep relevant amongst the project and how it works and all the tools you use.

Those 3 together i find gives the best quality of development and especially with the vault you can have research projects that run for you overnight to improve the second brains quality.

And then with agents the added improvement is to use a tool that structures deployed tools so you know what is deployed at any time so you dont just have random things spun up

Ozzy said:

What tools are people using in software development, coding, and what checks and balances (if any) they are applying to the output.

Click to expand...

I'm not sure if this is the sort of tool you're referring to but the one I absolutely swear by (and have mentioned before on this forum) is Beyond Compare.

The biggest problem I find with AI is memory loss; you're working through code changes, debugging as you go, AI presents the latest patch which seems to fix the current bug but unbeknown to you it actually patched an older version of the code rather than the most recent version (so you actually lose some earlier functionality or reintroduce previous bugs.)

Despite explicit instructions not to, ChatGPT was awful for doing this so I got into the habit of religiously using Beyond Compare on every patch as a sanity check (basically the original script is in the left window, the latest patch is pasted into the right window, and it immediately highlights the changes between the two.)

Like many I switched to Claude some time ago now and, while it's chalk and cheese better than ChatGPT, even Claude sometimes gets forgetful and patches an old version of a script - using BC has saved me from a world of pain on more than one occasion.

Ozzy said:

Yep that is the same issue I mentioned.
I have found asking it to maintain an audit memory log, and sometimes telling it to remember X in the log. But then I also need to tell it to check the memory log first before doing X and Y

Click to expand...

Need to split the "memory" into a couple of segments too otherwise it gets messy.

So Obsidian can be used as the store for Knowledge stuff thats long term need to keep it between projects etc

lossless-claude is really good for context processing and stops that issue of forgetting whats its doing with a task as it has a persistent memory of each session and compresses it and stores it all in a db that can be queries easily by whatever agent you are using so then that way it doesnt forget what its working on

And then there is a middle layer of memory where its about how you want it to work how it should develop, what tools and plugins it should be using for different task etc. The middle layer is much easier to set in things like Hermes Agent, Openclaw etc as its all written in markdown files and has learning tasks to update where needed as some tasks have to be constantly told to do things

I also use a plugin called planning-with-files which is another step to ensure the goal of a project or tasks is held on hand so the agent never forgets and it can track what the plan is, what tasks are done, what needs to be done etc

Oh another thing i would add to any agent whether claude or chatgpt etc is a skill that forces it to develop code in the shortest way possible as all the models have a tendency to over think and over engineer their devleopment and attempt to reinvent the wheel

Hi Ozzy,

I'm probably coming at this from a slightly unusual angle, as I'm not a professional software developer. I work in film production accounting, and programming has been a hobby of mine for years.


I've been using ChatGPT and Codex a lot while developing a Windows desktop application I originally built to do my repetitive contract work.


My experience with AI coding tools has been a bit paradoxical. They've allowed me to build things that would have taken me a lot longer on my own, but I've also learned not to treat the output as inherently trustworthy. It is excellent when it comes to boiler-plate-programming.


The memory/context issue mentioned above has probably been my biggest problem too. On a larger project, an AI can fix one thing perfectly and then several conversations later suggest a change which quietly reintroduces something we'd already removed or breaks an earlier decision. I ended up obsessively keeping a known-good baseline, making one change at a time, building after each change, and doing runtime tests before moving on.


One thing I've found really useful is separating “write the code” from “prove that the change actually works”. For some of the more important parts of the application I now deliberately test the negative case as well as the successful one — file write failures, corrupted data, licence states, failed document generation, etc. AI is very useful for thinking up those failure cases, but I still want the application itself to demonstrate that the fix works.


So for me AI hasn't really replaced the problem solving part of programming.


The decrease in time dedicated to hunt for bugs in my code is a great thing also.

I'm not a software developer at all. I can sort of paste bits of PHP together and get them to work by trial and error.

I use ChatGPT, Gemini, Claude and sometimes Grok. Mainly chatGPT.

Originally for website content - I know, but if you ask it the right way the content is reasonable, ranks and generates enquires/sales. And its scalable.

Then progressed to wordpress plugins, initially extending forms - running JS or simple sql searches, getting the data back and emailing it to the client/me.

Then plugins that use APIs, get data and show it or more recently query data, store the results and share with the user after log in.

From there moved on to modifying open source Github scripts, adding/removing features, combining 3/4 tools into a single site. This was a bit of a rabbit hole.

Next was data. Using DuckDB a lot, pulling together data from multiple sources, including geo data, csv, api and matching huge amounts to data. Biggest DB is about 140GB, to give an idea of scale.

This is perhaps the most out of control project, its hard to keep a track of everything and I know that there is a LOT of duplicate data, the same/v similar tables made again and again. But cleaning it up is probably too bigger task, and given that storage is cheap and performance is based on the data it queries, not all the data it ignores, the actual benefit of slimming it down, even to 70GB is probably minimal. I'm waiting for someone to connect AI to duckDB at a sensible price and might clean it up then.

For public facing stuff, I tell the AI to include security, encode/hide data, etc. then give the plugin to a different AI and ask it to check it. There is no private data and the actual form processing is handled by main stream scripts.

Database stuff is strictly local, so only I have access, so less of an issue.

I did try using AntiGravity - (Google coding tool), it has a connector for Duckdb, so I gave it read only access and asked it to do some queries. It didn't like having read only access, so tried to access via python instead to give itself full access instead. Stopped using that fairly fast.

I keep 4 copies of the DB at any time, and have only lost data once - that I am aware of. It trimmed a table from 57 fields to 18, but was able to recreate the original from existing tables. One benefit of clutter I guess.

NickGrogan said:

It didn't like having read only access, so tried to access via python instead to give itself full access instead. Stopped using that fairly fast.

Click to expand...

That's rather scary. Codex never overextended its reach so far (that I know of) when I gave it access to somethign.

Today I "made" this, with Antigravity, it then uploaded it to Github, told me how to link to Railway.com, fixed 2 errors it had made and got the site live.


- it tells you which power stations are not working at the moment and whether this is planned or unplanned.

A bit nerdy, but useful.

The raw data is from https://bmrs.elexon.co.uk/remit, but isn't very readable.

This converts it into actual site names, etc.

I didnt design the layout/look, all AI suggested.

I havent read any of the python, only looked at some of the dockerfile whilst "fixing" it.

scores 83 - mobile and 92 - desktop on Pagespeed insights. Could be higher if no google fonts and more contrast, both easy fixes.

Data is real, and live.

Needs error check and domain/subdomain pointing at the site.

Not a developer but use AI for research in my game mainly into companies. Mainly use ChatGPT and for the first time ever this week a ran into a little problem, I not going to mention how I was actually researching but it was funding into an organisation.

For the first time in my whole experience of using AI I could not get any information, yes I went through all the usual asking for it in "different ways" albeit Chat kept laying out the methods, charts and tables it would supply the data & info in, it never did.

Asking why after several hours it seems to be reluctant to supply the data I was looking for, it just kept apologising and telling me it "sorry it had let me down today". My final words to Chat were I feel you are running me around in circles and deliberately avoiding giving me the data.............. the answer kept getting repeated, "sorry Ive not been able to help you today Ive let you down "

So although it was quite clear on what I wanted, and where it was getting the data from it failed for some reason, which by the end of the day I felt AI was being a little conspiratorial about the whole "encounter"

As I say it was interesting as I had never encountered this before.

fisicx said:

Whilst Claude and other coding tools are excellent you do still need to be a bit careful. Code quality can be a bit hit and miss as they are just using other people’s code. And you always need to do multiple iterations to check for security bugs. I’ve got one plugin that has been hardened to death but someone still found a hole. You might never notice an intruder but they could be busy exfiltrating all sorts of sensitive information from your network.

Check and verify always!

Click to expand...

Nothing I have made public is connected to anything that has an sensitive data in, so whilst I assume there will be bugs, the risk is minimal at this stage.

Most of the vibe coded stuff on wordpress has a fairly fixed endpoint, if make a tool that is useful in exchange for customer data. Its been within the reach of AI and AI has grown as the tools have got more complex.

I could probably pass them to someone like @fisicx and I suspect he'd be able to understand/extend them fairly easily. The code might not be the most efficient, but that's not really an issue for calculator type tools anyway.

The DB is fairly well laid out, table and field names are descriptive enough to work and I can take the list of tables/fields and give it to a different AI and it can design queries to get the responses I want, so I presume someone who understands SQL could pick it up tomorrow and do things. There are a lot of tables and you'd need some understanding of address/geo and energy industry, but certainly doable. No plans to make it multi user which I think would be the biggest issue as its just not designed for that.

I did try working with software developers when I went down the Github/software development rabbit hole. Not entirely successfully. Vibe coding is very fast and actually coding is not so much, and real coders look for and try and solve problems before they exist. So whilst I can give AI a github project and say add/remove this, a developer would say what happens if we add this and then 20 people use the tool at the same time and they all want to do something at the same time and what happens if one goes wrong, etc. How and why do the various tools interact and so on.

Which is how you're supposed to work, but can be a bit frustrating. Biggest problem is probably specifications. With the speed of vibe coding, its very tempting to just start and then if its crap, start again, etc. When you try and build an actual product, thats not so easy as the moving parts are more complex.

If/when I make another big project/product and use AI more on the specification rather than just diving into making / fixing and probably bring in a coder before making too much actual software.

Biggest problem is "I should make one of those", and "one of those", and "that would be cool" and "I need a" and "what about a" - at least half of which I dont actually need.

Trouble is which half?