Kshitij Mishra | AI & Tech

I turned Claude into an entire company. 42 skills, organised like a real org chart (links below): Here is every department, and where to get each one. Developers Superpowers →github.com/obra/superpowe… Context7 → github.com/upstash/contex… Skill Creator → github.com/anthropics/ski… MCP Builder → github.com/anthropics/ski… Webapp Testing → github.com/anthropics/ski… Claude-Mem → github.com/thedotmack/cla… Designers UI UX Pro Max → github.com/nextlevelbuild… Taste →github.com/Leonxlnx/taste… Frontend Design → github.com/Leonxlnx/taste… Transitions → github.com/Jakubantalik/t… Web Artifacts →github.com/anthropics/ski… Brand Guidelines → github.com/anthropics/ski… Marketing 45 skills to run your marketing, from copywriting to SEO to lead magnets. Access them all here → github.com/coreyhaines31/… Social Media 17 skills to run your social media, from post writing to Reels to thumbnails. Access them all here → github.com/charlie947/soc… Finance 8 skills to run your finances, from statements to reconciliation to audits. Access them all here → claude.com/plugins/finance Small Business 31 skills to run your small business, from cash flow to payroll to invoicing. Access them all here → claude.com/plugins/small-… Legal 9 skills to handle your legal work, from contract review to NDAs to compliance. Access them all here →claude.com/plugins/legal Every skill on the chart is real and installable from the links above. Repost ♻️ this to help your network build their own.

@RodmanAi This feels like having a YouTube consultant for free.

Claude can now BUILD a YouTube channel from scratch and hit monetization in just 90 days. 100% free. Here are 8 prompts to launch your YouTube channel with Claude.

LLM fine-tuning techniques I'd learn if I were to customize them: Bookmark this. 1. LoRA 2. QLoRA 3. Prefix Tuning 4. Adapter Tuning 5. Instruction Tuning 6. P-Tuning 7. BitFit 8. Soft Prompts 9. RLHF 10. RLAIF 11. DPO (Direct Preference Optimization) 12. GRPO (Group Relative Policy Optimization) 13. RLAIF (RL with AI Feedback) 14. Multi-Task Fine-Tuning 15. Federated Fine-Tuning Since we're talking about fine-tuning, I wrote a full breakdown on fine-tuning LLMs with RL in 2026. Including how to skip manual reward engineering with automatic LLM-graded rewards. And this is done using a 100% open-source solution: github.com/openpipe/art

I'm sharing 8 of the best GitHub repos for automatically extracting data from any website: 1. Firecrawl Pass it a URL and it automatically crawls the entire site, converting it into clean data ready for AI. Supports JavaScript-heavy pages. Over 140,000 stars. github.com/firecrawl/fire… 2. Crawl4AI Turns any website into text that LLMs can read directly. No API key or payments required. Created by a developer fed up with services charging $16. Over 70,000 stars. github.com/unclecode/craw… 3. browser-use An AI agent that uses the browser like a human: clicks, logins, forms, everything. Created by a startup from ETH Zurich. Over 100,000 stars. github.com/browser-use/br… 4. Crawlee A complete framework for scraping with proxy rotation, retries, anti-detection, and queue management. Avoids blocks like a pro. github.com/apify/crawlee 5. Scrapy The classic Python framework for large-scale scraping. It's been in use for over 10 years and stably processes millions of pages. Completely free. github.com/scrapy/scrapy 6. Scrapling Automatically adapts to website design changes and avoids block detections. Paid-level features… but free. github.com/d4vinci/Scrapl… 7. AutoScraper Just give it an example of the data you want, and it learns the pattern on its own. Extracts the rest without writing selectors. Super simple and powerful. github.com/alirezamika/au… 8. curl-impersonate Imitates real Chrome browser HTTP requests to bypass the most advanced anti-bot protections. github.com/lwthiker/curl-… Save them all, highly recommended. With these, you can pull clean data from almost any website without paying for expensive services or relying on limited APIs.

@Bencera Proof that AI is becoming a force multiplier, not just another tool.

I grew Polsia from $0 to $10M run rate in 5 months. Solo + AI. Zero employees. Everyone asks me how. Presenting aisloP episode 2: “The Rise”

Most AI startups are selling AI. This one used its own AI to build itself. The founder of @Polsia just shared how they went from 0 to a $10M run rate in just 150 days. The interesting part isn't the revenue. It's that Polsia was already running parts of the company: → managing inbox chaos → helping with fundraising → reducing operational overhead → letting the founder focus on decisions instead of busy work The product became its own first customer. Worth watching if you're building with AI agents. This is not financial advice. Always do your own research.

I think this is something worth sharing — I tried TenPayGo on my recent trip to China. Signup was quick: just used my email (no Chinese phone number needed), got a verification code, and linked Apple Pay with one tap. No extra card entry. Then I was able to scan QR codes at restaurants and metro stations without fumbling for cash. One small thing: I used Gmail because I read that some other providers might get a “closed beta” popup — not sure, but it worked fine for me. Honestly, once it was set up, I didn't think about payments again. That alone made the trip feel smoother.

10 GITHUB REPOS THAT SHOULD BE ILLEGAL TO HAVE. all free. all open-source. bookmark this for later. 1️⃣ recordly — the free screen studio. open-source screen studio. auto-zoom, smooth cursor, webcam overlay, styled backgrounds, polished demos without an editor. (AGPL-3.0) 🔗: github.com/webadderallorg… 2️⃣ stirling-pdf — your entire pdf toolkit self-hosted. merge, split, sign, redact, OCR, convert, compress, 50+ tools, runs locally, nothing leaves your machine. (MIT) 🔗: github.com/Stirling-Tools… 3️⃣ photogimp — turns GIMP into photoshop. photoshop shortcuts, layout, and splash screen patched right on top of GIMP 3+. 🔗:github.com/Diolinux/Photo… 4️⃣ open notebook — self-hosted notebooklm. drop in pdfs, urls, youtube links — chat with them, summarize, even generate podcasts. bring your own model (18+ providers). (MIT) 🔗: github.com/lfnovo/open-no… 5️⃣odysseus — pewdiepie's self-hosted AI workspace. chat, agents, deep research, docs, email, memory — local-first, your hardware, your data. (MIT) 🔗: github.com/pewdiepie-arch… 6️⃣freedomain — free domain names for everyone. register a domain, point it at cloudflare or any DNS, ship your site without paying for the name. (AGPL-3.0) 🔗: github.com/DigitalPlatDev… 7️⃣ hyperframes — write HTML, render video. heygen's engine that turns html/css + animations into deterministic mp4s. built for AI agents. (Apache-2.0) 🔗:github.com/heygen-com/hyp… 8️⃣ web-to-app — turn any website into an android app, on-device. configurable webview, apk signing, even node/php/python runtimes — no remote build. 🔗: github.com/shiaho777/web-… 9️⃣ reclip — self-hosted video + audio downloader. paste a link from youtube, tiktok, x, ig — 1000+ sites — grab it as mp4 or mp3. powered by yt-dlp. 🔗: github.com/averygan/reclip 🔟 excalidraw — the infinite whiteboard that replaces miro, figjam and lucidchart. hand-drawn diagrams, wireframes, real-time collab, end-to-end encrypted. 120k+ stars for a reason. (MIT) 🔗: github.com/excalidraw/exc… most people pay monthly for tools that already exist for free. you don't have to.

Stanford researchers did it again. They just introduced what might become **Git for AI agents.** Here's why that's a big deal. Long-running AI agents don't just generate text. They create files, install dependencies, launch servers, update databases, fill KV caches, and build an entire execution environment. Imagine the agent is deep into a task. Everything works perfectly... Until one bad decision. Maybe it misinterprets an error message. Maybe it overwrites the wrong file. Now the tests fail, and the agent spends the next several minutes trying to repair damage it created itself. Every failed attempt means: • More tokens consumed • More tool calls • Bigger context windows • More expensive runs The usual fixes aren't great. You either: → let the agent keep digging a deeper hole or → restart everything from scratch. Restarting sounds simple, but it means paying for all the earlier reasoning again. Even worse, agent runs aren't deterministic, so the second run often takes a completely different path. The missing piece has always been state. Traditional agent logs only remember conversations and tool outputs. They don't preserve everything happening behind the scenes. Things like: • running processes • installed packages • temporary files • memory state • open handles • KV cache Git only tracks files. It can't restore an executing environment. That's where **Shepherd** changes things. Instead of recording a plain chat history, it captures the execution as structured runtime events. Each step becomes a snapshot of the *entire* environment—not just the source code. That means you can jump back to an earlier checkpoint and continue from the exact same running state. No rebuilding. No reinstalling. No repeating the prompt. Because the earlier context stays identical, Shepherd can reuse over **95% of the KV cache**, dramatically reducing replay costs. Its copy-on-write snapshots are also reported to be around **5× faster** than Docker commits. The coolest part? You can place another AI agent on top as a supervisor. If it notices the primary agent drifting toward a bad solution, it simply rolls back to the last healthy checkpoint and lets it continue from there. Almost like hitting **Ctrl + Z** for autonomous agents. There are limits, of course. Rolling back local files is easy. Undoing an email that's already been sent... or reversing a payment... isn't. Those actions still need safeguards before they're executed. Stanford evaluated Shepherd on public benchmarks. On CooperBench, adding a supervising agent increased collaborative coding success from **28.8% to 54.7%**. It's still in alpha, but the concept is incredibly promising. Git transformed software development by making code changes reversible. Shepherd is trying to bring that same idea to **AI agent execution**. Repo: github.com/shepherd-agent… ⭐ Worth starring if you're building AI agents.

10 GitHub repo for trading on Polymarket all of this is free and can help automate your trading 1. prediction-market-backtesting: github.com/evan-kolberg/p… backtest trading strategies on real Polymarket and Kalshi data 2. TradingAgents: github.com/TauricResearch… lets you build a complete trading system 3. /last30days: github.com/mvanhorn/last3… analyzes news from the past 30 days 4. Assistant Tool: github.com/FiatFiorino/po… indicators that show where the market is heading 5. Firecrawl: github.com/firecrawl/fire… turns any website into clean data 6. Pydantic AI: github.com/pydantic/pydan… agent framework for building production bots 7. n8n: github.com/n8n-io/n8n news analysis and filtering 8. Tavily MCP Server: github.com/tavily-ai/tavi… search built 9. Collector & Analyzer: github.com/txbabaxyz/coll… collects the full trade history of any wallet 10. Binance Collector: github.com/txbabaxyz/mlmo… predicts market direction and calculates fair value completely free

@Shruti_0810 Foresight Reasoning is a clever systems design.

Everyone keeps trying to turn video generation models into robot brains. LingBot-VA 2.0 takes the opposite approach. It was built for robot control from day one—jointly modeling future visual states and robot actions instead of bolting control onto a video model afterward. The details are what make it interesting: → Semantic visual-action tokenizer → Sparse MoE for efficient inference → Native causal video-action pretraining → Foresight Reasoning for real-time closed-loop control → Few-shot adaptation and in-context learning Feels less like "AI for robots" and more like an architecture designed around how robots actually interact with the physical world.

@Shruti_0810 Realistic practice beats memorizing scripts every time.

One thing AI is surprisingly good at is creating realistic environments to practice. A customer conversation isn't a script. People interrupt, change their mind, get frustrated, and ask unexpected questions. Training should reflect that reality. This is a smart direction.

Wait... this completely changed how I use Claude. For months I thought better prompts were the answer. Then I found LOOPS.md by Andrej Karpathy. It's not another prompt template. It's a framework for making AI work through problems instead of giving you a one-shot response. After trying it, I realized I'd been using only a fraction of what these models can actually do.

Live customer conversations are no place for agents to learn on the job. Cresta Training Simulator replaces scripted role-play with live, dynamic AI simulations that help human agents practice before going live. Learn more ➡️ cresta.com/blog/cresta-la…

@cresta The biggest shift here is moving training from static scripts to realistic conversations. That's much closer to what agents actually face on day one.

@hamidteghra Crazy results 👀

Github repo link : github.com/averygan/reclip

SOMEONE JUST OPEN-SOURCED A TOOL THAT LETS YOU DOWNLOAD VIDEOS FROM 1000+ WEBSITES FOR FREE. Meet ReClip. A lightweight, self-hosted video downloader with zero ads, zero trackers, and zero paywalls. Drop in a link from YouTube, TikTok, Instagram, X, or hundreds of other platforms... Choose MP4 or MP3... Download in seconds. Here's what you get: → Works with 1000+ supported websites → Download videos as MP4 or audio as MP3 → Pick your preferred video resolution → Download multiple links in one batch → Modern browser-based interface → Easy self-hosted deployment The growth has been wild. 1.4K+ GitHub stars in just 9 days. 239 forks and climbing. Why is everyone starring it? Because people are tired of sketchy downloader websites filled with popups, fake buttons, and tracking scripts. ReClip keeps everything on your own machine. No accounts. No limits. No data collection. Just a fast, open-source downloader that does exactly what it's supposed to. MIT licensed. Built with HTML. Only two contributors. 100% Open Source. Free forever. 👇

@Shruti_0810 The world model approach is what stood out to me too. Excited to see how it evolves.

I think we've been measuring video models with the wrong question. Instead of asking "Does it look realistic?" Maybe we should ask: "Can it predict what should happen next?" That's what caught my attention in LingBot-Video. It combines: Sparse MoE (30B total, 3B active) 70,000+ hours of embodied robot data Reward models that score physical plausibility, motion, and task completion—not just aesthetics. The benchmark numbers are impressive, but the bigger idea is treating a video model as part of a robot's world model rather than a content generator. Curious to see where this direction goes. @robbyant_brain Model: huggingface.co/collections/ro… Code: github.com/Robbyant/lingb… #ad

@RodmanAi Already saved this. Can't wait to go through each prompt and see which one fits my goals best.

CLAUDE CAN BUILD YOU A PASSIVE INCOME MACHINE IN JUST ONE WEEKEND. Here are 10 prompts ready to copy and paste. Save this thread before everyone else does.