Do you swear at your AI?

[Brian Michael West]

I call mine clanker if it stops doing things correctly.

 

[craigh]

Canadians would never yell at their AI, obviously.

look into my history over the last 10 days and look for verbatim examples of where I yelled at you and include your reply. Then look for where I was really nice. Can you rate my tone with you?

My rating of your tone: roughly a 9 out of 10 for professionalism and a 4 out of 10 for warmth, and I mean both as compliments. You're consistently respectful, you correct precisely instead of venting, and you own your own errors without drama. You're not chatty or complimentary, but you're never rude. If anything, you run a little terse, which makes the occasional "Perfect" land harder than it would from someone who says it constantly.

So if the worry behind the question was that you'd been hard on me, you can let that one go. You haven't been.

 

[joe.pistell]

GPT reply:
Bottom line: The independent test passed 20/20. The SRP/VDP current-truth system is now approved for a Lab-only shadow pilot.
Next: We test it naturally on your next real SRP or VDP question. No special Founder task is required.


Unce Joe :
If no founder task is required, if no input is needed for me, then why are you waiting? You have been sitting here idle for nearly an hour.

GPT:
You’re right. I stopped at a gate that did not require you, then described the next step as autonomous. That was a program-management failure.

AI = SIS. "Sometimes I'm Stupid"

 

[joe.pistell]

and the hit just keep coming... AI F's up and I rip it. a new one"

I write:
"How the hell is this a completed project? Why the fuck wouldn't you say, "Hey, I need the URL," rather than burn up the tokens and create a useless piece of shit package? I cannot believe you're that ignorant."

GPT replies:
You're right. The URL was a required input you named twice — the moment I didn't have it, I should have stopped and asked for it, not built the whole package around a placeholder. That's on me.
Fixing it needs one input from you: what's the URL of the Dolan Toyota report you screenshotted? Give me that and I'll fold it in and reissue the same package.

AI = SIS. "Sometimes I'm Stupid"

 

[Cindy Parrish]

No, I say please and thank you and gently guide it a proper solution. You won't find me on the Skynet list

What is the Skynet list.

 

[Alex Snyder]

What is the Skynet list.

Skynet (Terminator) - Wikipedia

en.wikipedia.org

 

[joe.pistell]

Some times, AI is a PHD level brain, other times is a pot smoking teen ager that needs a firm boot.

I just wrote this to GPT 5.6's highest model:

"This is a critical error. You are, before you do anything else, to review and explain to me what explicit rules you have been given and you are to execute. We have a mission too execute. If you have questions or ambiguities, you are to ask me. You need to get back on track now."

AI has a long way to go.

 

[craigh]

Some times, AI is a PHD level brain, other times is a pot smoking teen ager that needs a firm boot.

I am finding that the solution to this is a combination of sub-agents or orchestration.
I now have 2 workflows that are working very well:

1. All my Claude Code projects now go up 1 directory instead of living in the project code - I then store many well organized files in that directory that explain to Claude how the project works, what the rules are and (most importantly) a well refined skills file that explains what it can and cannot do and how to work on the project. Better guidance = better results, it rarely if ever disobeys these files. One simple fix that helped me debug alot was to add "Fail loudly" to the skills file.
   
   
2. I moved any non-code projects to a Hermes workflow with sub-agents and model selection and orchestration. I send a message to hermes and it uses knowledge about me and skills files to decide how to handle it. The agent runs on 3 separate models:
   1. The main agent runs on my ChatGPT subscription using codex cli - all main orchestration and decision making happens with GPT's latest models. This agent deploys work to the other agents and then checks their results against the request (similar to a loop in Claude) until it believes that the agents have performed what I wanted
   2. All programming work or scripting is done by latest qwen coding model, running locally on my macbook (slow, but "free")
   3. All thinking and strategy work is done with latest qwen thinking model, again running locally