Srinivas Rao

Skills were broken on claude.ai for 5 days last week. During the outage Claude Code wrote this script, we edited it in a document editor it built, and uploaded it to YouTube through an uploader it built. None of that required the thing that was broken. @AnthropicAI

@bcherny @teja2495 Skill update flow is brutal — to change anything you have to move files to tmp, edit, rezip, then manually re-upload and replace. Skills should be editable in place. The whole point is persistent behavior but the update cycle kills that entirely.

@teja2495 Curious, what bugs are you hitting on desktop? Would love a short list

With the Claude desktop app being very buggy for coding, I switched completely to the Claude Code CLI. It is excellent. I now understand why many people prefer the CLI over the desktop app.

@emollick You have specifically tell it to read your skill file so it uses that instead. Feel free to DM i've built dozens of them. Loved Co-Intelligence BTW, also a Portfolio Author.

Is there any way to turn off the frontend design skill in Claude for the web? It seems like some default Anthropic skills are always available to the chatbot, making it hard to get it to use your skills when doing tasks the default skills also cover.

Tokenmaxxing is leaving the engine running in your Lamborghini so people in the parking lot know you have one- Rich people setting money on fire and calling it innovation.

But even more important, the ability to read and go deep on anything, gives you the ability to break AI’s frame and apply what you’ve learned. Reading your Taste for Maker’s essay this morning sparked and that led to multiple insights about how we should design for AI - same concepts are still relevant.

The fact that AIs tend to answer you in bulleted lists tells us something important, though somewhat depressing: people can't read. They don't do this by accident. What you're seeing is an implicit portrait of the median user.

@kevinroose Tokenmaxxing" — Silicon Valley really is just vibes and vocabulary at this point

One OpenAI employee used 210 billion tokens *last week*. A single Claude Code user spent $150,000 in a month. Meta, Shopify and other companies now factor token use into performance reviews. Are you tokenmaxxing, anon?

I wrote about tokenmaxxing, the new status game taking over Silicon Valley. nytimes.com/2026/03/20/tec…

@atShruti wait a founder showed up to a meeting with you without a deck? WTF? lol

Its week before YC demo day and time to complain about valuations, uncapped notes, LOIs, no deck 1 zoom meeting investment process, founder fraud, and the weather! It’s tooo hot in SF!!

@atShruti And it forces accounability both ways. If you wanted nothing other than investor who just writes a check, that isn't much different than a loan from a bank.

@atShruti Exactly the kind of investor, I'd want, one that wants to meet more often not less.

A founder once told me, "The best investors don't just wire money - they tell you what you need to hear, not what you want to hear." That only works when founders trust you're in their corner. When an investor challenges your assumptions, be grateful. It means they see your potential. • They want your company to win. • They care about you as a founder • They see what you're building before the market does. Great investors hold you accountable because they know that's how companies go from good to category-defining.

anyone else having issues wiht Claude "This isn't working right now. You can try again later." I was able to start a few therad, but they all collapsed in like a few turns.

@atShruti Just reading that has my mind spinning with a world of possbilities.

What a world we are in. You can run & fine tune billion parameter models on phones now with Tether's new BitNet framework. No dependency on cloud or NVIDIA. Lot of defense applications! I remember I couldn't even deploy a small app with some compute on my Nokia years ago.

@atShruti also dm'd you with al ink to try the product for youreslf, literally write a task description, press a button, get the result. I can give you specific on tasks to try via dm if you want, but it's pretty self explanatory considering it's fill out a form and push a button :)

@UnmistakableCEO That's really cool, will check the email.

Token Spend to overtake cloud spend this year? AI costs going up despite the layoffs. We are backing companies that can help here and interested in more! Reach out if you are building here!

the interviewer background has given me a massive edge I had no idea would be so useful. My rule of thumb is always assume the first thing an llm tells you is bullshit and question the assumptions. Since it operates from training data , while humans operate from lived experience, no contest :)

@UnmistakableCEO It’s an incredible time to be alive… the key is the ability to be curious and express what you want.

Watched a friend demo his new app yesterday. Built the whole thing in a weekend. No code. Just prompts, an AI agent, and a clear vision of what the customer needed. Two years ago that was a $50K MVP and three months of dev time. The barrier to entry didn't lower. It evaporated. 🌊

From 9 months of using Claude Code, I have learned that the more you leave open to interpretation in an AI task, the more time you spend cleaning up the mess. Ambiguity is not flexibility. It is a bug you scheduled in advance. Constraints are not limitations — they are the only reason anything ships clean.

@HarshSh54928171 Biggest tip: keep skills narrow and load them conditionally. Every skill injected on every call eats context window whether you need it or not. One focused skill per task domain beats a giant swiss army knife. I'm writing a non-dev guide to Claude Code that covers this in depth — DM me if you want early access.

Writing "The Non-Developer's Guide to Claude Code." What's your biggest question or frustration with it? Drop it below — might end up in the book.

It's like a highway that seems fine at 2am. Then rush hour hits and you realize it was never designed for actual traffic. The road didn't get worse. The volume exposed what was always a shitty road.

The agents are basically a stress test that every existing tool fails instantly. They amplify the exact weakness that was always there — the human was the coordination layer, and the coordination layer was always the bottleneck. We just didn't notice because the creation rate was slow enough to mask it.

AI agents don't solve the knowledge management problem — they expose how broken it always was. When a human creates 3 notes a day and can't find them in a week, that's a personal discipline issue. When agents create 50 notes a day and nobody can find anything by Friday, that's an architectural failure.