What I find that works well, an MCP to a central sync platform that can move emails between accounts, create filters. Thin clients for any device. AI is connected to my project platform (twiglit) and my notes app. Also the agent can make changes as it needs. Also I have a token system so any agent can get shortlived access (read only or write or delete) which is account specific. AI gives a daily overview and tells me what filters or created.
My friend, RTM! This is 100% on you. The help docs say clearly that it does not mask metadata, nor your profile and custom styles.
It does not grant anonymity from Anthropic or the model. It’s being quite transparent about your lack of understanding what it means by incognito.
“Can Claude access my profile information (custom styles, personal preferences, etc.) in incognito chats?
Yes, Claude can access this information within an incognito chat.”
https://t.co/4GeAyopmo1
Just give this time, you’ll change your mind.
1/ use it enough and it’s super annoying style of prose will begin to grate on you.
2/ it’s independence is great in some cases, but as you want it to do things more consistently, it will get frustrating.
3/ the liberties it takes can be catastrophic at worse, but often a big time drain fixing its overreach
It does have a higher level of intelligence, but I think it’s oriented to weaker users.
Let us know how it goes in 6 months.
Yeah it’s absolutely the wrong pattern of thinking that people have. They are thinking of it like a person when it’s more like electricity.
Humans as well are just meat bags, but with sparks of intelligence (if only occasionally).
A computer is useful because of the selected application of continuous electricity.
We simply now have on-demand judgement. The idea of an agent is largely a distraction. What’s important is the combination of selected use of both electricity and judgement.
Do you even use AI? What on earth are you talking about? If AI is causing technical debt it’s because you aren’t asking it to correct those problems. Do you really tell yourself that Anthropic, ChatGPT, Uber, Cloudflare, SpaceX, and every other large tech company are using AI as jr developers and ballooning technical debt?! Total delusion.
@theshawwn Dude, just take up either indoor climbing or golf. Meeting guys is really not hard. Or like the infinite number of meetups around every subject.
@jasonfreedman Yeah, it’s a hard rule for me doe AI: you are not me, do not write for me. Suggest ideas, give feedback, only write for me when explicitly told to.
@HarperSCarroll The industry will get there. AI is just a different form of electricity. At some point it will be the same, developers should think of it as such.
@ivanburazin@daytonaio Building Meant, it makes AI-built software
decidable, authorizable, and provable. Hook it to any conversation and it will implement what you decide and give you a link to check the work.
I don't give the AI much leeway in what it is asked to do. It has tasks in Twiglit that are always very concrete actions.
Affiliate Program Application for Parachute (task container)
- Apply to Parachute on https://t.co/pJJJF9wivl
- Log the application result (in Twiglit)
- Review the Parachute application result (in the email client)
Most of the work is like this. It is more likely to make a mess duplicating task containers, but there is no real excuse for deleting or moving things.
A long time ago when I first saw (on here) AI doing rogue actions, put command intercepts in place to catch really destructive commands. I also add to the various configs a second layer of restrictions.
Within Meant, my IDE, AI no longer has access to the repo trunk, for sure not prod, and lives in short runs in the cloud.
I know it's a waste of tokens, I know it will probably get me in trouble with the ai overloads, but I can't resist giving my librarian an an insult generator tool for the response when an ai passes a certain level of error and garbage in their commits.