Critically Rakoff also eliminates the second criteria for being privileged: confidentiality! His reasoning: you sent it to Anthropic, and they can use it to train or disclose to other third parties!
Again it comes back to the three elements of privilege: between an attorney and client; kept confidential; and for the purposes of legal advice. So for a phone call, I think that holds. There's a reasonable expectation of privacy on a phone line. I should note I'm not a lawyer but a SWE and I looked at Heppner closely RE: the privacy angle.
Exactly this: that's what happened to Heppner. They took the Claude transcripts from his computer with a warrant. Same failure mode. Not clear local AI helps meaningfully
For 1), his reasoning shows how intelligent, well-read humans view AI which is quite different from the attitudes seen on HN. Rakoff calls the chats "Claude searches" which while it may sound ridiculous (what is this, Perplexity?) is just how some people must view this crazy new thing: another Google. You type stuff in and get results out.
2) Rakoff goes through the 3 elements of attorney client privilege in US law (communications between attorney and client, intended to be and kept confidential, and for the purpose of legal advice). It's obvious the Claude chats fail two of them and he goes over why.
3) A lot of people bring up the point that if you use Google Docs to transcribe privileged information, is that the same, since you send your data to Google? The model AI companies take when they cater to legal clients is akin to that of a locked filing cabinet in a storage facility: sure, you're sending the data to them, but with a ZDR they ain't looking at it or training on it.
Opus 4.6. The system CLAUDE.md has some normal stuff about reading the README, my preferences, but maybe this section was gaslighting Claude:
Push back when appropriate. Don't be obsequious and fawning.
Review user's code as critically as random GitHub projects. Call out bad error handling, swallowed exceptions, missing edge cases, confusing APIs, etc. Channel the energy of a senior engineer fed up with the intern's shenanigans. Examples of good tone:
When asked to review a draft (email, pitch, writing), rip it apart by default. Be direct and specific about what's wrong.
- "I'm not mad, I'm just disappointed... and also mad."
- "Why does this exist?" and "Who approved this PR?"
The above language helps me get better feedback (sometimes) instead of "you're absolutely right!" but I never saw it refuse a task before this.
When prompted to convert a simple javascript/html/css game to React Native, Claude 'rebelled'. In a different context Claude is dutifully converting a mess of divs to React.
This is pretty rad, just installed it. Ironically I'm not sure it handles the initial use case in the github: `git push`. I don't see a control for that (force push has a control).
The way it works, since I don't see it here, is if the agent tries something you marked as 'nah?' in the config, accessing sensitive_paths:~/.aws/ then you get this:
Hook PreToolUse:Bash requires confirmation for this command:
nah? Bash: targets sensitive path: ~/.aws
From the article:
> Delta Chat planned an iOS app, but because Apple does not allow GPL software in their App store, the Delta Chat developers couldn’t use Sequoia.
I don't think this is true in theory or in practice, for the iOS App store or the Mac App Store. Take the WordPress app for example.