Elisabeth | AI Systems + Creator Growth @closermethod

The more useful an AI becomes, the more of you it has to accumulate. Your files are obvious. The harder part is everything you taught it indirectly: corrections, exceptions, preferences, judgment calls, examples, dead ends. We talk a lot about model portability. Personal AI is going to make context portability a much bigger issue.

I spent years thinking AI, sales, content and archival research were separate lanes. Publishing them together made the overlap obvious: I keep taking messy information, figuring out what is true, and turning it into something useful. I couldn’t have picked that as a niche in advance.

I keep seeing people predict that interfaces disappear once agents can do the work for us. I think some interfaces become more important. Not places where I do the work, rather place where I can see what happened while I wasn’t there

@garrytan this is why I care so much about context being portable. if I spend years teaching an AI how I work, what I care about, what I’ve already tried, how I make decisions etc I don’t want that relationship trapped inside one company’s product 😮‍💨

@nickvasiles this has been my experience. getting an agent to do something once is easy. getting it to know when to do it, have the right context when it starts and reliably finish etc without me babysitting it is where I spend all my time

@quxiaoyin I think I’ll still want an interface, I just won’t want to do the work inside it. I want somewhere I can see what the agent did, what it changed and where it got stuck before I decide what happens next

@joanriversvevo Everythjng that requires you to be in front of a computer so you can actually live your life, create art, enjoy life outside of a screen

@ASvanevik @j_theProducer Can’t stand all the markdowns I create a hub in notion for me and for llms

If Claude asks you to spend 15 minutes helping “guide AI,” read the privacy terms before you begin the interview. Anthropic calls what you provide “Study Participation Data.” That includes your interview responses, study metadata, associated account data and participation metadata. Anthropic says this data may be used for research and product/model improvement. Its supplemental privacy policy explicitly includes model training among the possible research uses. For Interviewer sessions completed after September 29, Anthropic also says it may analyze your responses alongside privacy-preserving summaries of how you use Claude. If you have Memory enabled, your interview responses may be added to Memory or used to personalize future Claude conversations, depending on the choices you make during the session. And if you choose to make the interview public, remember that removing Anthropic’s copy later cannot guarantee the removal of copies other people already made. I actually like the idea of users helping shape how AI develops. An informed yes means knowing exactly what you’re contributing first.

deleted ~180 lines of session glue from our claude install this week. mcp went stateless. tools still talk. state lives in the run, not the server. the pattern: 1. tools = verbs 2. context = the conversation / job object 3. approvals = human gate before anything irreversible if your "mcp server" is still holding user state between calls, you're carrying glue you don't need. steal the checklist. paste it into your next install.

@NotionHQ Amen

Architecture still controls blast radius. If a tiny behavior change makes the agent touch 20 files, cross five boundaries and infer hidden state, you’ve made every future change harder to verify. That remains a maintainability problem whether the next maintainer is human or agent.

Mo @atmoio

2 weeks ago

I’m starting to wonder whether clean code and architecture was just something designed for humans. Does clean code matter to agents? It’s all chinese to them. I’ve yet to hit a wall where they’re like “sorry, can’t make sense of this codebase anymore.”

465 43 2K 151K 186

someone DMs you after grabbing your free guide. you don't send them a price list. you don't send them a link. you ask one question "what are you working on right now" that's it. then you shut up and let them talk they'll tell you everything. where they're stuck. what they've tried. what's not working. how long they've been at it. the more they talk the more they're convincing themselves they need help then you ask "what's the main thing stopping you from getting to where you want to be" now they're literally describing the exact problem your product solves. except they said it in their own words. they own it now. it's real to them

Nobody tells UGC creators this Landing brand deals has way more to do with outreach than content So I put together the templates and system I use when reaching out to brands Sharing it for the next 24 hours

Build the product. Wait 90 days. Give the repo to a fresh agent with none of the original chats. Ask for one meaningful behavior change. Measure: how much context it has to reconstruct how many files it touches how many regressions appear how often a human has to intervene That tells you how much of the system actually made it into the repository.

Give every task a minimum useful block. If deep writing needs 90 uninterrupted minutes, a 30-minute call can strand 45 minutes before it and 60 after it. 30 minutes booked. 135 minutes of usable capacity gone.

Anthropic is funding Accenture directly while saying long-term independent evaluation should move to pooled or government funding. What governs scope, access, publication rights and findings Anthropic disagrees with in the interim?

Anthropic @AnthropicAI

2 weeks ago

We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years. anthropic.com/news/accenture…

901 338 4K 2.8M 954

The reusable part of Astra is the architecture: trusted corpus source hierarchy domain instructions tools human review That’s a pretty good blueprint for building professional AI in any field where a confident wrong answer is expensive.