Ravikant Dewangan

Last weekend I hosted Greg Glassman at my gym for a 2-day @MetFixByBSI seminar. What I learned broke my brain. I went down a rabbit hole I couldn't escape and when I finally came up for air, I felt like I'd been unplugged from the Matrix. Now I'm building a course to teach what I learned. 🧵👇

@IPStrategyGuide @Codie_Sanchez That is exactly why domain expertise matters. You can not just trust what comes out of the model. You need someone who knows the field to validate the output and catch what the model gets wrong.

@ronitkd @Codie_Sanchez General model will be full of inaccuracies due to models pulling from and training on non-true source data.

Claude for legal work is saving me 6 figures already this year... (also it's just as good as Harvey shockingly).

@galdo_galdo_ Separate workspaces for each project. I keep a dedicated context doc for each one so I can switch without losing the thread. The real trick is not mixing them in your head.

@ronitkd how's context management working across 12 projects? i get tangled switching between even 3

12 apps on one whiteboard. 3 monitors running Claude Code sessions in parallel. This is what shipping looks like when AI handles the scaffolding and you handle the taste. No venture funding. No team of 50. Just one person who refuses to stop building.

@IAmAaronWill That notes step is key. Most people skip it and go straight to prompting. Writing it out first forces you to clarify your own thinking. The AI just amplifies whatever clarity you bring to it.

@ronitkd Too true. My process is literally write it down in notes, then feed it to Claude

Don't learn AI. Build with it instead. • Pick your biggest time sink • Open Claude Code • Map the workflow around it • Identify every manual step • Build the automation in n8n • Add your MCP servers • Connect it to your tools • Test it end to end • Fix what breaks • Document the system • Find the next bottleneck • Repeat Your competitors are watching tutorials. You're shipping infrastructure.

@HarakiriInu That is the only way. Build something, break it, learn from it. Repeat.

@ronitkd doing by learning 👍

hot take: you don't need to know how to code to build software anymore I built vidno with claude code and cursor the skill is knowing what to build and how to instruct the AI

As a CrossFit coach, this is the one concept I wish more people understood. Your body has two fuel tanks. Most people only use one. New article breaking it all down.

Built a metabolic health calculator this week for my coaching clients. Pulls fasting glucose and insulin from their bloodwork. Outputs HOMA-IR and TG/HDL ratio. Flags insulin resistance before symptoms show up. Claude Code. One afternoon. Would have taken a developer 2 weeks and $3K. edu.metfixseattle.com/tool-metabolic…. .

Built a metabolic health calculator this week for my coaching clients. Pulls fasting glucose and insulin from their bloodwork. Outputs HOMA-IR and TG/HDL ratio. Flags insulin resistance before symptoms show up. Claude Code. One afternoon. Would have taken a developer 2 weeks and $3K

Glassman was right about one thing most fitness people still get wrong: sickness, wellness, and fitness exist on a single continuum. Your blood work tells the story before your body does. We built @MetFixByBSI around this idea. Stop guessing. Measure your metabolic markers. Train accordingly. The people who treat health like a data problem outperform everyone running blind. More on this framework: edu.metfixseattle.com

Seeing this play out in real-time shipping SaaS apps. The CLAUDE.md file is now the most important file in our repos it's the spec, the architecture doc, and the prompt layer all in one. Structured markdown is how you program agents. The developers who figure this out first have a massive edge.

Markdown is the new Python

@cyrilXBT Can confirm. We went from prototyping to shipping 5k+ revenue apps with Claude Code at Autosterea. The learning curve is steep for about a week, then it clicks and you realize you've been doing 10x the work manually. Best investment of 2 hours anyone in tech can make right now.

You have 2 hours tonight. You could watch Netflix. Or you could learn Claude Code and come out the other side able to build real AI apps. Your call. Full course 👇

@esteban The security testing use case is underrated. Most teams can't afford dedicated pentesters on every release. Having an AI that can chain together attack vectors across a large codebase in an hour changes the economics completely.

For the record using Opus 4.6 it took about an hour with limited tokens to build few successful e2e attack chains in a large multi million LOCs code base. In one case in black box testing it was able to find a vulnerability and iterate for hours successfully. This is game changer

This is the insight most agent builders are sleeping on. The model is already good enough — the constraint is what it can see. We've seen the same pattern building production agents: pipe in the right domain knowledge and the agent goes from toy demo to replacing entire workflows. Context is the new code.

karpathy autoresearch experiment improves results 3.2% over baseline by simply giving the agent access to ML papers give your AI agents access to knowledge they are superhuman in literature search

@ifourth Exactly this. I replaced 3 separate tools with Claude Code automations in my agency. Saved about $400/month and the workflows are faster. Start with what already costs you money. That's where the ROI is immediate.

If you have an existing business, the way to make more money with AI isn't building something new. As a solopreneur, here are the 2 highest ROI moves you can make right now: 1. Use AI to replace things you're already paying for Especially monthly subscriptions. If you're an ecom seller, build your own simple Shopify app instead of paying for one. If you're a SaaS founder, use AI to migrate your SaaS to Hetzner. This alone saved me $25,000+ a year. That's like adding $2k MRR without selling a single thing. 2. Use AI to automate your daily workflow and replace employees For me as a POD seller: creating designs, uploading products, launching ads. As a SaaS founder: building content systems to promote the product. Don't build something new until your existing business is lean and running on autopilot.

The agents that win this year won't be the smartest ones. They'll be the ones with the best context pipelines. We're building this at @Autosterea right now — the model is table stakes. What separates production agents from demos is how cleanly you feed them the right data at the right time. Most "AI failures" are actually context failures.

@AnandButani Exactly right. The bottleneck shifted from "can you build it" to "should you build it and does it actually solve a real problem well." Taste is pattern recognition from years of seeing what works and what doesn't. The models give you leverage but taste is the multiplier.

The taste part is what nobody talks about. That's the actual job now. With a wide enough monitor or enough monitors, you can build a ton of apps. These models are powerful

@liamtrn They won't replace developers. They replace the boring parts so you can focus on the parts that actually need a brain. The trick is reviewing everything like you wrote it yourself. That part doesn't go away.

GitHub Copilot suggested I use a deprecated function. Claude told me to import a library that doesn't exist. ChatGPT generated a SQL injection vulnerability. And somehow these AI tools are going to replace developers. Meanwhile I'm debugging code written by an AI that can't debug its own suggestions.

@IAmAaronWill That's the move. The notes step forces you to think through the logic before the AI does. Most people skip straight to prompting and wonder why the output is mid. Your workflow is basically spec-driven development — and that's exactly how we ship with Claude Code at Autosterea.