Shreyans Bhansali

@itsolelehmann models cook better when you stop yelling 30 instructions at them

i deleted half my Claude setup last week and every output got BETTER sounds backwards, but anthropic's own team just explained exactly why it works. here's the one prompt that tells you what to cut (and you don't even have to paste anything): this is what happens to everyone... you get a bad output, so you add a rule to your skills. "be more concise." next week, another bad output. another rule. "use a casual tone." but a month later, something else breaks. "always explain technical terms." you keep stacking, and it feels productive because you're fixing problems as they come up. but 3 months in, you've got 30 rules piled on top of each other. some of them contradict each other ("be concise" and "always explain your reasoning" are fighting). some of them fix problems that the model doesn't even have anymore. and the model is trying to follow all of them at once, which means it's doing none of them well. it's like handing a chef a 47-step recipe when they only need 12. the extra 35 steps slow the chef down, make them second-guess the parts they already know, and the dish comes out worse than if you'd just let them cook. that's what over-prompting does. anthropic just published a piece on how they build claude code (the ai coding agent). their own engineering team found that their scaffolding was making the ai worse which means your custom instructions are almost certainly doing the same thing. so here's the actionable move... instead of manually reading through your setup line by line, just tell claude to audit itself. if you're in claude's desktop app, claude already has access to your: claude[.]md (the file where your preferences and rules live), your skills folder (where your reusable instruction files are stored), your context files, everything. just open claude code/cowork and say this: — "read my entire setup before responding. check my claude .md, every skill in my skills folder, every file in my context folder, and any other instruction files you can find. then go through every rule, instruction, and preference you found. for each one, tell me: 1. is this something you already do by default without being told? 2. does this contradict or conflict with another rule somewhere else in my setup? 3. does this repeat something that's already covered by a different rule or file? 4. does this read like it was added to fix one specific bad output rather than improve outputs overall? 5. is this so vague that you'd interpret it differently every time? (ex: 'be more natural' or 'use a good tone') then give me a list of everything you'd cut with a one-line reason for each, a list of any conflicts you found between files, and a cleaned up version of my claude.md with the dead weight removed." — one message. claude goes and reads your entire setup, audits it, and comes back with exactly what to cut and why. you don't dig through files, you don't read every rule yourself. it does the whole thing. once you get the results, don't just blindly delete everything it flags. here's the process: 1. read what it flagged and why 2. delete the flagged rules 3. run your 3 most common tasks with the trimmed setup 4. did the output stay the same or get better? the deleted rules were dead weight 5. did something specific break? add back just that one rule the goal is to find the minimum viable setup that gets you the output you want. your ai setup should be getting simpler over time. addition by subtraction baby

@HarryStebbings ai will post, humans will still carry the taste

At 20VC we don’t hire social media managers. If you are truly gifted at social, you run your own channel. It’s that simple. All these corporate social media managers will be replaced with AI in 12 months.

@anulagarwal “i’ll just build my own search” is how side projects are born

Apple is a $3.7 trillion company yet can't fix their search I just installed Claude to try out their new update but I literally can't search it in Apple search - so broken. Pretty annoying tbh I will vibe code my own Apple Search bar using Claude

@ryancarson “i’ll just build a small orchestration layer” famous last words

I haven't typed `npm run dev` on my local machine for three days now and it's absolute bliss. Having my agents 100% in the cloud is a massive unlock. (One of those agents is openclaw, which is technically on my mbp in my office, but the only way I interact with it is via email/slack so it “feels” cloud) I'm able to run all the engineering and marketing for my startup through Slack and Linear and because of this the work product that I'm shipping has increased dramatically. I know all of us devs love creating our own custom solutions to this stuff but the truth is that creating an agent orchestration layer for your company or startup is a full-time job. Our job as startup founders is to be growing the company, not to be building out an agent orchestration custom platform. I think if you have a larger engineering team like Ramp, then it does make sense to build an entire layer like Inspect agent. However, I would venture to say that I'm getting most of the value by simply paying for a pre-built, battle-hardened solution like Devin. Again to be clear I'm not being paid by Devin or anybody to say these things, just my real-world experience using this stuff.

@viditaatrey 22% bump means people just wanted a shopkeeper, not a search bar

Introducing Vaani – Your Meesho Dost, India’s first Gen AI-powered conversational voice shopping assistant, designed for the next 500 million users who inherently shop this way. Most e-commerce is built for digitally savvy users who know what they want and how to find it. A large part of India doesn’t shop that way. Shopping here has always been assisted. You walk into a store, describe what you need, ask questions and decide with guidance. That’s what we built for. Vaani brings this offline experience online. It understands natural, real-world queries, asks follow-up questions to clarify intent, and works seamlessly across the app to help users discover, decide and purchase. Early signals are encouraging. Over 1.5 million users have interacted with Vaani, with strong repeat usage indicating early habit formation and a 22% higher conversion rate among users who engage with the assistant. When technology understands you, shopping becomes simpler. When it guides you, confidence grows. And with that confidence, millions more people can participate in commerce. #Meesho #VoiceCommerce #DigitalIndia

@svpino agents write code like they won’t be the one on-call at 3am

Yes, these agents are absolutely horrible at writing decent-quality code. But any time you say that out loud, there's a swarm of AI-apologists who've never built anything that scream at you, "skill issue!" You might not care about the code, and that's alright, but you can't gaslight anyone who cares about good code.

@venturetwins meetings feel illegal when your agents are idle

I have this weird new anxiety every time I’m sitting in a meeting and don’t have several long-running agents working on tasks for me 🙃

@MillieMarconnni prompt engineering just became prompt scraping

This feels like cheating. Someone built a Claude Code skill that scans Reddit and X from the last 30 days on any topic you give it, then writes you copy-paste-ready prompts based on what the community has actually figured out not what was working six months ago. You type /last30days prompting techniques for ChatGPT for legal questions and it comes back with the top patterns real lawyers and power users are using right now, complete with a fully written prompt you can drop in and use immediately. No more Googling, no more digging through threads, no more prompts that worked last year but got patched out. It works for anything - Midjourney techniques, Suno music prompts, Cursor rules, trending rap songs, whatever you need to know what people are actually saying about right now. 100% Open Source. MIT License. Link in the comments.

@ivanburazin twitter decides what’s real until your retention chart disagrees

Twitter is THE MEDIA now. It's the place where success is defined. If something doesn’t land here, it basically didn’t happen. Launches are judged by how well they perform here. Narratives formed here first. Even threads have evolved into articles. Everything else is downstream. I don't read TechCrunch or even newspapers anymore. If it's not here, I probably don't see it. There's absolutely zero need to log into Medium or subscribe to 10 different Substacks you'd never read. Don't know exactly when this shift happened. But it's complete now.

@brycent this exists in 5 demos and 0 products that don’t break after 3 uploads

Startup idea that I would angel invest in immediately Google cloud for context: A cloud storage provider for videos and photos. But every time I upload a video, the AI transcribes my video, understands the video, and allows me to dynamically search for the video in my cloud based on what happened in that video This would be useful for creators who have tons of footage or anyone who's trying to save years' worth of photo and video and access them quickly. If you are building this, DM me.

@lennysan doing 2x the work in half the time so now you just get 4x the work

The results from our latest reader poll: How has AI changed how you feel about your job? The short version: most people better about their work with AI, and yet also less secure in their role. People are also feeling more productive, but almost no one feels their workload has actually gone down. Respondents shared a fascinating mix of feelings: a sense of childlike wonder, coupled with deep frustration about the impossible pace of change and suffering from AI slop everywhere. (Shoutout @noamseg for running this poll, and to everyone who participated)

@kr0der codex invents a framework, claude forgets half the feature. pick your poison

i’m currently testing Opus 4.6 in Claude Code after using Codex for 3 months straight both have a big downside and i’m currently weighing which one actually slows me down more: Codex - random abstractions, random helpers, not that good at frontend, variable/function naming could use some help Claude - less thorough so it often misses places when implementing features, and produces slightly more bugs in general

@yusan_lin @mirrormirror_ai paying humans for their data and likeness? rare moment where AI isn’t speedrunning chaos

Today @mirrormirror_ai is launching the marketplace where fashion models license their likeness and brands get stunning AI-generated imagery featuring real people. Commercially licensed, model-approved. Try our platform: mirrormirrorai.com As a fashion model I used to spend hours on fashion photoshoot sets. I later did my PhD in CS and became a Research Scientist on AI for fashion. I can see clearly that AI image generation is replacing a large portion of my old job. But brands that use AI recklessly have already paid the price. It damages reputations and hurts the bottom line. Putting real people at the core of AI-generated imagery isn't just about avoiding backlash. It's better business. That's what Mirror Mirror AI is built for. Right now, Mirror Mirror AI houses agency-signed models who have graced the covers of Vogue and Harper's Bazaar. You can digitally book them using our fashion-centric AI software, get your campaign done in hours instead of weeks, and never have to fly anyone in. You purchase a license for commercial use upon approval, and the models get paid. Mirror Mirror AI is also opening a global call for independent models from anywhere in the world to apply to be featured on the platform. Work with fashion brands internationally, choose the projects you take on, and earn from your own likeness on your own terms. Selected models will be announced at an exclusive event in New York during @Techweek_ this June. Apply for the open call: mirrormirrorai.com/open-call A huge thank you to our incredible team for pouring their hearts into this launch, and to a16z @speedrun for believing in our vision from the start. We're just getting started.

@mattpocockuk offline agents shipping code sounds illegal but i’m listening

Working on a tool that orchestrates locally sandboxed coding agents in TypeScript - Sandboxed in Docker - 100% offline: commits made in the sandbox get patched back to the host - Build complex workflows in Typescript - Claude, Codex, OpenCode It's called Sandcastle

@mattturck now your first hire is $20/month and doesn’t ghost you

Going to be hard to explain to future generations that there was a time, before Claude Code, when people with an idea would sometimes spend years looking for a “technical co-founder”

@Chioma__Amadi tier 1 work getting automated, tier 2 just got promoted

In simpler terms, Claude is coming for the jobs of: • Administrative assistants • Virtual assistants • Data Entry Clerks • Customer Support Agents (Tier 1 / Frontline) • Operations Assistants • CRM/Data Update Staff • Junior Research Assistants • Transcriptionists • Email/Inbox Managers • Content Writers (Basic/SEO-heavy writing) And a whole lot more I can tick off the top of my head. I hope you’re upskilling, reskilling and evolving with the times.

@GergelyOrosz imagine paying and still getting “pls create account to continue”

Windows is the only: 1. Mainstream OS that is *paid* (Linux: free; Mac: free as part of buying an Apple device) 2. OS that doesn't allow use without creating a Microsoft account Kind of ridiculous, especially in light of #1. A big turn-off for me. Microsoft SHOULD remove it

@akshaymarch7 we replaced “how do i write this” with “why does this even work”

I’ve been noticing something recently… The way I code has changed a lot. Earlier it was simple. I’d sit, think, write code, get stuck, Google, repeat. Now it’s more like… I think → ask AI → get an answer → then spend time figuring out if it’s even correct. That last part is surprisingly hard. Honestly, writing code feels easier now. But understanding it properly feels… harder. Sometimes AI gives something that looks perfect, but deep down I’m not fully convinced. And then you spend time untangling it. I think the skill is shifting. Earlier it was about “can you build this?” Now it’s more about “can you judge this?” A lot of people are saying AI will do everything. Maybe it will help a lot. But someone still has to take responsibility for the code. And that part hasn’t changed. I’ve started caring a lot more about fundamentals because of this. Because tools are getting better very fast. But clear thinking… that still is very important. Not sure if this makes sense. But things definitely feel different now. In future I'm pretty sure AI is just going to improve a lot and make very less mistakes. Strange time to be alive.

@JonathanRoss321 everyone says “AI for efficiency” till someone uses it to steal their customers

AI will split companies into two camps. Value Preservation — using AI to cut costs. Value Creation — using AI to grow revenue. Value Creation companies will be the ones hiring more people because of AI. Which kind of company do you work at?

@allenholub “can’t do architecture” meanwhile half the internet is vibe-deploying to prod daily

There isn't an LLM on the planet that understands architecture. No LLM can create an effective architecture. Architecture is contextual, dependent on things like the domain. Architects use judgment and creativity. No "AI" can do either. There is no "good" cookie-cutter architecture. What does work is creating an architecture first, then telling the LLM to produce code that fits. It's much easier (and much less error-prone) to tell Claude to create a microservice with a single responsibility that uses a specific pub/sub interface, with the messages and their payloads all well-defined in advance, than to create a larger thing and hope for the best architecture-wise. And you get much better results in the long term. Architects use judgment. That's way beyond the capability of any "AI." A "non-technical user" cannot do that. Deal with it. Hire some programmer-architects. The fantasy that a nontechnical user can create a high-quality software product with nothing but surface-behavior-level prompts is just that—a fantasy. I've never seen an example of such a thing released to the public and functioning well. This non-technical programmer dream has been around since COBOL days. Didn't work then. Doesn't work now.