LiquidMetal AI

Hi, I'm Basia. I haven't opened ChatGPT in 8 months. This is either addiction or great business: I run my whole company on one AI tool now. No, Anthropic isn't paying me (I checked) I just kept replacing workflows - writing, research, design, automation, code - and every single one ended up at Claude. (I'm on every Anthropic product I can find, and I've vibecoded my own agents, built custom skills, designed my infographics, run my budget in Claude , etc. A therapist would call this a problem.) Here's what I've replaced in 12 months: 1️⃣ Writing & strategy: ChatGPT → Claude ↳ Holds the thread across 50-page docs without losing the plot. ↳ Will tell you when your idea is weak instead of cheering you on. 2️⃣ Research & search: Perplexity → Claude (Research mode) ↳ Pulls citations, structures the brief, and writes it back to you. ↳ One pass instead of 14 browser tabs. 3️⃣ Design & decks: Canva + Gamma → Claude (Artifacts + Skills) ↳ Infographics, one-pagers, slides — all generated inside the chat. ↳ I built a brand skill that ships my visuals in my own style every time. 4️⃣ Automation & workflows: Zapier → Claude (Skills + Agent SDK) ↳ Skills give Claude reusable capabilities. Agents run them on a schedule. ↳ My weekly content pipeline runs as one agent. No more wiring up zaps. 5️⃣ Building apps & coding: Cursor + Bubble + Lovable → Claude Code ↳ Builds production apps from a description. Ships, iterates, debugs. ↳ I stopped outsourcing builds and just shipped the tools myself. That's 8 tools out, one vendor in. My SaaS bill dropped, my output got faster, and the tab clutter finally died. Either I've stumbled into the best operator setup of 2026... Or I need to stage an intervention on myself. (I'm not staging an intervention.) Which workflow would you move to Claude first? Drop a comment below. ♻️ Repost to help operators in your network simplify their stack. And follow Basia Kubicka for more on building a solo business on one AI stack.

This is how to turn Claude into your private Tutor 👇 Steal these 6 prompts: Most "learn with Claude" advice is useless: "Build me a 20-hour plan." "Make me a cheat sheet." "Quiz me until I break." You're treating Claude like a textbook generator. But that's not what makes it powerful... The shift: stop asking Claude to feed you content. Start asking it to pressure-test your thinking. I've spent the last year learning new skills this way - almost daily. And it's the single biggest unlock I've found. Here are the 6 prompts I actually use: (See the visual for the full breakdown 👇) 1️⃣ Diagnose before you teach ↳ "Before you teach me [topic], ask me 5 questions to figure out what I actually know, what I think I know but don't, and where my blind spots are. Then build the session around those gaps." 2️⃣ Translate into my world ↳ "I'm already deep in [my field]. Explain [topic] using only analogies from that world. And tell me exactly where each analogy breaks." 3️⃣ Teach me the wrong mental models first ↳ "Don't explain [topic]. Tell me the 5 most common ways smart people misunderstand it. Then show me why the wrong model feels intuitive." 4️⃣ Stress-test, don't quiz ↳ "Don't quiz me on definitions. Give me 3 real-world scenarios where the textbook answer is wrong, incomplete, or contested." 5️⃣ Project first, theory second ↳ "Give me a small real project that uses [topic]. Don't teach me anything first. When I get stuck, teach me only what I need to move forward." 6️⃣ Map the territory ↳ "Give me the 30,000-foot map of [topic]: what's settled, what's contested, what's just trendy, and where serious practitioners disagree." The pattern: Every prompt makes YOU do the cognitive work. Claude becomes a sparring partner, not a content vending machine. That's where the speed actually comes from. Which of these 6 will you try first? Drop a comment below. ♻️ Repost to help curious PMs in your network. And follow me ( Basia Kubicka ) for more on building with Claude.

7 steps to build AI Agents from scratch even if you’ve never done it before. I've spent the last few months building agents with Claude Code. And the same 7 steps that used to take me weeks now take an afternoon. (Save this before you build your first one 👇) 1️⃣ Define your use case ↳ One problem. One success metric. Clear boundaries. ↳ Bad scoping = broken agent. Always. 2️⃣ Pick your scaffolding ↳ LangChain, CrewAI, AutoGen, LlamaIndex... or Claude Code. ↳ Pick on complexity, not hype. 3️⃣ Connect your LLM ↳ Claude, GPT, Gemini. Configure params. Set keys. Test responses. ↳ Your agent is only as smart as its brain. 4️⃣ Add tools ↳ Web search. Email. DB queries. APIs. ↳ No tools = no action = just a chatbot. 5️⃣ Add memory ↳ Short-term for the conversation. Vector DB for the long game. ↳ Without it, your agent forgets everything after one turn. 6️⃣ Add a planner ↳ ReAct. Plan-and-Execute. Tree of Thoughts. ↳ This is what controls how your agent thinks and loops. 7️⃣ Add guardrails ↳ Rate limits. Input/output filtering. Human-in-the-loop. Evals. ↳ Skip these and your agent does something you can't explain. Then test, deploy, monitor, improve. That's the full loop. Most beginners skip steps 5, 6 and 7. Then wonder why their agent hallucinates, loops forever, or burns $300 in API calls in one weekend. That's exactly why I built my upcoming LinkedIn Learning course: Vibecoding Your AI Agent with Claude Code. It walks you through it step by step Drops soon - I'll share the link the moment it's live. Which step do you think most builders skip first? Drop it in the comments below. ♻️ Repost to help builders in your network ship their first agent. And follow Basia Kubicka for more on building AI with Claude Code.