Pratik Agarwal

तो क्रांति हो तुम जब मौके को फ़र्ज़ मानते हो, तो क्रांति हो तुम। जब दर्द से डरते नहीं, बल्कि उसे अपना रास्ता बनाते हो, तो क्रांति हो तुम। जब तन-मन को औज़ार में गढ़ते हो, तो क्रांति हो तुम। जब हमसफ़रों में वही जज़्बा जगाते हो, तो क्रांति हो तुम। जब ऐसा जीवन चुनते हो, जिसकी तपस्या से जगत उठता है, तो क्रांति हो तुम। wrote for my father on his b'day today..sharing here for all Kraantikaari founders ❤️

Claude Code: Review Permissions If you use Claude Code, you've probably noticed it asks for permission a LOT — especially for Bash commands. I built a skill that fixes this. /review-permissions scans your entire Claude Code history, finds every command you've been approving over and over, and generates the exact settings.json rules to auto-allow them. It even groups suggestions by confidence level (safe read-only stuff vs. things worth reviewing). What it does: - Analyzes all your past sessions across every project - Shows you how much time you've spent waiting on permission prompts - Identifies which commands aren't covered by your current settings - Proposes rules sorted by frequency, lets you pick which to apply - Edits your settings.json for you To use it: 1. Drop review-permissions.md into ~/.claude/commands/ 2. Open Claude Code in any project 3. Run /review-permissions That's it. Takes ~30 seconds and you'll never approve git status again. If you want the MD.. just reply and I'll DM!

🤔

A child who creates something every day will always outperform a child who consumes lessons every day.

Mark Manson said it right: “Beware: learning more is a smart person’s favorite form of procrastination.”

@TheKatareKid haha interesting example.. obviously LLMs can do much better.. and ofc I appreciate the brain is very complex. but I think we're not very far from markdowns that express specific judgement/intuition

x.com/mazeincoding/s… Was going to write a long comment on how this is extremely wrong. But this quote tweet does a better job 😅 Simple answer - Brains can’t be replaced by LLMs ever, let alone .md It’s technologically and mathematically not possible. Lot of current jobs can be replaced yes, but not brains. The humbling thing should be the low level quality of work we have been doing, which can be predicted by a pretty dumb brute force prediction machine, that works because the system was scaled to unprecedented levels in hope that AGI will happen. It won’t, but big tech/ big ai will still call anything that is good enough to replace jobs as AGI. 🤷🏻‍♂️

It’s humbling to accept that we’re heading to reducing human brain to a few markdown files 🤯

@pktrpl but is it really.. even our intuition in specific areas can be conveyed to the LLMs via some markdown.. such as how we respond to an email or something more abstract like how we like to draft a blog?

@_pagarwal I think of it more like reducing the memory of human brain down to a few hundred markdown files, and the comprehension/reasoning skills have been replicated into a neuralnet

@thericebowlgirl I’d love to read that!

@_pagarwal no we’re not - it’s not that simple. i think i’ll have to do a post on why not, and why this is just crazy simplification.

@lego_sharat Good point!!

@_pagarwal A few markdown files and lots of GPUs & CPUs :). Always amazed by how energy efficient our brain and body is.

Build agents or work for them. Only two options.

@Konstantine x.com/_pagarwal/stat…

Product roadmap multiples as you release it!

YC is the real Lambda school!

APIfication of humans begin!

Four types of people at every company now yes, people get 10x better when the go from bottom right to top right but also, people get 10x worse when they go from bottom left to top left

Most founders think they’re “using Claude.” They’re not. They’re using the chat window. Claude Code in the terminal is a different species. It’s not autocomplete. It’s not Q&A. It’s an AI employee. The branding misses this completely. If it were called “Claude AI Employee,” adoption would 10x. The gap between using Claude and using Claude Code is enormous. Try it once. You’ll realize you weren’t using Claude at all.

This might be the most important framing shift I've read on agents. The "human-in-the-loop" vs "human-as-governor" distinction cuts through so much confused product design. Most teams are building agents that ask permission at every decision point—which just recreates the coordination bottleneck with extra steps. The real unlock isn't agents that can do more tasks. It's figuring out which decisions humans actually need to make vs which ones we're making out of habit or fear. What breaks first: our ability to build agents that can handle autonomy, or our willingness to give it to them?

@abhiroopm Absolutely that will be a great idea. I’ll play along w this idea of task relevant maturity! Keep you posted!

Great post! Humans will indeed become the rate limiting factor. Makes me think - to really make this happen, we need Andy Grove’s Task Relevant Maturity framework for AI agents. Letting go of control to an AI agent is best done gradually, based upon its past experience and demonstrated ability to do similar tasks flawlessly.

@andrewchen x.com/_pagarwal/stat…

anyone with theories on how software teams will evolve? today: 1 eng, 1 PM, 1 designer (EPD) tomorrow: 10 PMs who vibe code all day + 1 eng architect (the architect creates the scaffolding and writes adversarial agents to manage tech debt, security, scalability issues) Thoughts?