Chat GPT & Generative AI tools

[firlandsfarm]

I’d also keep one short “current state” document in the project with the important decisions, assumptions and anything that must not change.

100%
"Give me a summary of where we are so far that I can paste into a new clean thread. " can be useful.

Yeah I use that (especially when Claude tells me it's run out of Tokens for the conversation! ). I have one, a legal matter with a supplier (I won! ) ,,, it ran into 5 conversations!

[Trader724]

I’d probably keep it much simpler. Once a conversation clearly starts drifting into a different subject, start a new chat rather than letting one thread grow into everything.
Then for any chat worth keeping finish it with a short factual summary and a few keywords. That should make it much easier to find again later without having to build a complicated system around it.
Not perfect but probably easier to stick with in practice.

[ShaunWhite]

Wow Shaun, I had no idea those controls existed and influenced. So you don't put them in the prompt, they go in the Project's 'Sources'?

They were the results from some long employee disciplinary meetings. "How can improve my responses" stuff, by default it's tuned to lightweight meaningless chit chat (fast and light on compute)

I've got it in my settings>personalisation>custom instructions but even it admits it doesn't look at them when you start a session.

So if i feel things drifting i'll stick them in a message and remind it to stick to them.

[firlandsfarm]

Wow Shaun, I had no idea those controls existed and influenced. So you don't put them in the prompt, they go in the Project's 'Sources'?

They were the results from some long employee disciplinary meetings. "How can improve my responses" stuff, by default it's tuned to lightweight meaningless chit chat (fast and light on compute)

I've got it in my settings>personalisation>custom instructions but even it admits it doesn't look at them when you start a session.

So if i feel things drifting i'll stick them in a message and remind it to stick to them.

OK, thanks...

[Big Bad Barney]

My understanding is that you do better to always starting new contexts/conversations. Keep them short. AI has bias towards the start and end of a conversation so it loses what's in the middle to some degree......Human's I believe have the same bias....but my understanding is....biased

[sniffer66]

Using GPT SOL 5.6, and with memory turned on, I find just giving it the name of the old chat, and asking it to update itself from there is enough to start afresh with all memory retained now. It didn't used to but it will even retrieve any code from the previous. I used to have to re-upload it. As son as it starts to get a bit laggy I start anew

[rosecruz]

I recommend looking into Obsidian and setting it up as a permanent BRAIN in your working folder (it really is much easier than it might sound below).

You will need to work locally. So for ChatGPT you will need to install Codex, and for Claude you will need to install Claude Code.

With Obsidian on your system, you can create a way of working by which all tasks within the project are saved both as findings and in a kind of journal/diary. So as long as you instruct the AI tool to always save what you work on, a permanent, persistent memory is created. I have set up reference manuals for all my databases. Those reference manuals have DB structure, column names, and explanations of what each is. Obsidian at the same time maintains an index of everything and somehow interconnects all the documents by linking them.

If I say "remember when we looked at the winning strike rate of jockeys that wear pink socks @Chepstow..." the AI tool will find those results for me.

Finally, I have a to-do list, so if I think on the fly about something I might want to do after a task is completed, I instruct the AI tool to add it to the list. Also, as you work on queries, more avenues to explore pop into your head, so again you add them to the to-do list. This guarantees that when you shut down your PC, the next time you start the AI tool, it will remember what you worked on, what the major findings were, and what the to-do list says.

Every time I start Claude Code, I paste this into the chat:

Ok we are now working on the xxxx project located at
C:\Users\H\Desktop\xxxx
Read its CLAUDE.md and follow the START HERE section fully before doing anything else, then confirm you're ready for queries.
Also look at the project's brain to understand better what we have been working on recently and the overall structure of this project

===================================================

TOP TIP: so you NEVER lose this BRAIN, create a GitHub (it's free for 1GB) account and a project, and instruct the AI tool to create a daily process to push the files created onto GitHub. If your PC fails, you will have a GitHub repository of all documents, all findings, and the whole structure of the project BRAIN. GitHub does not allow you to save DBs, so you will need a different backup process for the DBs. I personally back up to a second drive in my PC, and for some key DBs I have a process that updates to Backblaze, which gives you 10GB of free storage.

If you get lost, just explain to the AI tool what you want to achieve with Obsidian and that you wish to implement it with best practice, and it will do it all for you.

One final piece of advice: always limit the project to the folder, and activate either Time Machine on Apple or File History on Windows to back up daily to an external drive on your PC. This might sound over the top, but there have been instances where the AI tool has gone rogue and deleted whole DBs. This is because models are becoming more agentic, meaning you present them with an idea/task and they just launch forward to accomplish it, so we need to stay alert and have backup processes that are outside the AI tool environment.

Hope this helps someone

[Trader724]

I’d add one more precaution to all of that: if you’re letting an agent work locally I’d keep it sandboxed to the project folder and give it as little access outside that folder as possible.

I also prefer having it ask for explicit confirmation before it creates, overwrites, moves or deletes files, and certainly before it runs anything destructive. Reading and analysing can be fairly open but changing the filesystem should be a deliberate step.

Backups are obviously essential but ideally you want two layers of protection, stop the agent wandering somewhere it shouldn’t in the first place then have Git/File History/etc. as the safety net if something still goes wrong.

As these tools become more agentic distinction becomes increasingly important. “You can work on this project” shouldn’t automatically mean “you have permission to do anything you like on my machine.”

[firlandsfarm]

Using GPT SOL 5.6, and with memory turned on, I find just giving it the name of the old chat, and asking it to update itself from there is enough to start afresh with all memory retained now. It didn't used to but it will even retrieve any code from the previous. I used to have to re-upload it. As son as it starts to get a bit laggy I start anew

Sniffer, I used to (try to) refer to other chats in my GPT Plus account but it always responded that it cannot access other conversations. But I've noticed this before where it gives different responses in different chats. I had a good example with Claude. Because of it's irritating system of limiting the number of Tokens consumed in a conversation I tried asking if it could monitor the %age of tokens consumed and warn me when it reached 75% and 90%. It said yes, and for other chats just say that is what I want. So for my next new conversation I did, it said it couldn't! I asked why not, it said because it cannot access that information, I said you did before and pasted that part of the first conversation and it basically said I was lucky, some portals can access and some can't!!

[firlandsfarm]

I recommend looking into Obsidian and setting it up as a permanent BRAIN in your working folder (it really is much easier than it might sound below).

Thanks for the tipoff rosecruz, I've asked my developer to consider what you say rather than build our own database. I particularly like the use of 'plain English' to interrogate the set-up.

I’d add one more precaution to all of that ...

Actually more than one Trader but all good points.

[Trader724]

I recommend looking into Obsidian and setting it up as a permanent BRAIN in your working folder (it really is much easier than it might sound below).

Thanks for the tipoff rosecruz, I've asked my developer to consider what you say rather than build our own database. I particularly like the use of 'plain English' to interrogate the set-up.

I’d add one more precaution to all of that ...

Actually more than one Trader but all good points.

I was grouping them together a bit too neatly. Definitely more than two layers once you start looking at it properly.

[firlandsfarm]

I was grouping them together a bit too neatly. Definitely more than two layers once you start looking at it properly.

I confess to not having had the time to explore how I could use Git for anything through my lack of knowledge, not it's lack of ability but would it just be acting as a safe back-up in which case is it different to one-drive and others? I use one-drive solely as a home to back-up my HD ... all my files are on my home device.

[Trader724]

Git is much more precise for working on projects. You can see exactly which files and even which lines have changed, roll back just one particular change, compare versions and create deliberate checkpoints before letting an AI agent do anything risky.

Put very simply OneDrive is more like “give me the file as it was yesterday”, whereas Git is “show me exactly what the agent changed and undo just that bit”.
For code or projects with lots of text files Git is much more useful.