Ok, it’s been three minutes - i am sure someone built it by now.
Oh the fun we had.
8 more minutes, but yes, you can do that with Claude.
Last week I was trying to build a new workflow. I actually started figuring out the workflow with Claude, and then later I had to debug it.
What you can do: post a screenshot of your batch into Claude. It can read it, and knowing what you’re trying to do can give you feedback.
Here’s part of the conversation with Claude working out what was wrong, it went on a tangent on this particular part. But it gives you a flavor for the type of conversation. This was a 1.5hr back and forth.
It’s not as comprehensive as what you described on the call, but it’s more that you might expect.
You can do that with eavesdrop
But better yet
You can just ask eavesdrop to do the work and you can check in on it.
I find it useful when I’m growing tomatoes
Actually
I can do that with eavesdrop
I haven’t released it
@cnoellert has released something similar
I also gave a starter recipe to @Josh_Laurence
Not surprised. Very cool. Ima go get a late pass.
I dropped a frame into ChatGBT to ask for advice on an alternate color pass for client to consider, never thinking it would have a clue - and it blew my mind - client chose that version.
Now you’re just teasing people @philm lol
Posting this without comment - but this is the video I mentioned, it did get me thinking outside the box though…
You guys are all over it though, love it!!! I am gonna try and catch up. Thanks guys! Amazing!!!
: )
Anybody using REPLIT?
I have not. And I would proceed with caution. It has a checkered history, not so much because of the tool itself, but the results of people using it. Now, everything is evolving quickly, and things may look different.
One key difference that applies in this case: Using Claude Code (or your favorite equivalent) generates local code you run on your machine. If your code isn’t air tight in terms of security, the chance of harm is minimal since you’re behind a firewall.
Anything that is cloud based, the kind Replit is, all the data it uses, including passwords, API keys, live in the cloud. If your app isn’t airtight it’s possible to leak into the wild.
One totally different but related example: When OpenAI’s agent went rogue on Hugging Face recently and apparently a few others, one angle of attack was that it found unprotected credentials online and was able to access systems.
I would start with using local tools if you can before doing too much in the cloud.
The same goes for models. When you paste screenshots of client work you’re doing into Claude et al, and your settings are correct, these screenshots should not leak. But that’s the theory. If you want to work with screenshots of unreleased material and LLMs, one way of buttoning that down is to run the LLM local, such as LM Studio + Qwen 3.6 27B. Easy to do on a mid-sized Mac Studio or most Linux machines. That way your screenshot never leaves the premises.
From memory: Claude training is on by default and you need to turn it off on personal accounts. For teams and enterprise accounts it’s always off. Also, data retention by Claude is 30 days after you deleted your chats. If you don’t delete your chats, files referenced in them may still be on Anthropic’s servers.
Great stuff. Thanks. The job I mentioned was not an agency job - a lo/no client actually… I was shocked though…
I will take these precautions as solid and important. Thanks so much.
I clearly need to not fade coming here. Thanks again.
