One hotkey on my Mac pins the screen at 100% brightness and blocks sleep. Another reverts it. Claude Code made building both easy.

One snag: macOS auto-brightness kept dimming the screen, from 100% to 59% in ten seconds. Claude Code tracked down the hidden setting that switches it off.

This is a golden era for personal, malleable software: you shape tools to how you work instead of waiting for someone to ship them.

A short, succinct podcast on one of the biggest achievements of the Modi government: eradicating Left Naxal terror from India.

Bharatvaarta: How India Is Ending Naxal Terror | The Untold Strategy Behind the Final Crackdown FT. Tuhin Sinha Episode webpage: https://www.bharatvaarta.in/

Media file: https://traffic.megaphone.fm/ISP9790010887.mp3

Open source software is a cheat code for AI agents.

Recently, I was trying to paste rich text into a tool, and the tool kept messing it up. I asked an AI agent what was going wrong. It looked at the source of the tool, figured out how it handled copied data in the text area, and gave me formatted text that I could copy into the text area while retaining the format.

Pretty soon, we will design our spaces to be robot-friendly.

@innate_bot: Today, we are open-sourcing the Innate OS, the intuitive agentic OS for general-purpose robots 🎮 Try it today in our simulator and make the smartest robot ⬇️ Here is a demo with two robots collaborating We are entering the physical agents era, but building for mobile robots is still reserved for PhDs. We designed the Innate OS for intuitiveness, so that everyone can start automating their lives. Innate OS runs on our <$1k open-source robot or on your computer. It takes just a minute to build your first agent in our simulator.

I’ve stopped deciding about new tools from their marketing pages alone.

Claude Code already knows my setup: my dotfiles, my hooks, the rules for how work gets planned and launched. So when a tool shows up, I ask one question. Claude goes through everything about the tool (its website, its GitHub repo, its issues, its code) and checks all of it against how I work. I don’t read any of it myself.

This week that was Paseo, an open-source app for running Claude Code sessions from your phone, browser or desktop. Could it replace my home-grown setup for running sessions in parallel?

Claude told me exactly what it could replace in my workflow today: juggling terminal tabs to keep track of parallel sessions, and reaching those sessions from my phone.

It also told me what it couldn’t: the machine that runs the sessions, the planning rules every piece of work goes through, the way I set up a fresh checkout for each task, and everything I’ve built outside session management.

Both lists, checked against my setup, made the decision for me.

An evaluation that used to take an afternoon of reading is now fast and automated. It’s also personalized, which no feature comparison is. The question was never “is this tool good”. It was “does it fit what I already run, and make it better”.

Founder-led. Founder-less. Wartime CEO. Peacetime CEO. Managers are crucial. Managers are bloat. Micromanage. Be hands-off. Strict OKRs. Fluid goals. Tech is the differentiator. Tech is a commodity.

Throw a dart at any of these extremes and you’ll hit a highly successful company that swears by it.

Pick what works for your reality, not what sounds good on a podcast.

@davidsenra: “We’re not big on founders.” “I know this is maybe ironic to say given your podcast.” “We try to eliminate the idea of a founder from this company as much as possible. We think it distracts people from the company. The company is the center and whether you’re a founder or you joined a little bit later, all that matters is your contribution, your trajectory.” — Bending Spoons CEO Luca Ferrari

I rent a small server that never sleeps and run coding agents on it. I installed Tailscale for one reason: I didn’t want a box sitting on the internet with ports open. It quietly took over three other jobs I hadn’t planned for.

The first is the obvious one. Nothing on that machine is reachable from the internet — not SSH, not the database, nothing. Scan it and there’s nothing to try. Everything I do goes over Tailscale.

Then the keys stopped mattering. Server, laptop, and phone all talk to each other over Tailscale SSH and I don’t copy a key anywhere. The device is already proven when it joins. I’d been rotating keys by hand before that and not enjoying it.

The one I didn’t see coming: plenty of sites just block datacenter IPs. So the agent’s browser goes out through my laptop or my phone at home, as an exit node. Runs on the server, looks like my couch. And if that path drops, the browsing stops — I’d rather it fail than quietly go out the datacenter IP.

Last one is my favourite. An agent hits a login screen at 11pm and gets stuck. One tap on my phone and I’m looking at the server’s desktop. I sign in, hand it back. No port forwarding, no VPN app, no jump box.

Four unrelated problems, all solved by Tailscale.

@apenwarr: That’s the problem with Tailscale, nobody ever talks about us. Best kept secret. Why are you all keeping your friends in the dark like this

Another day, another evaluation.

This time: terminal-browser, a real Chromium browser that draws inside your terminal.

Claude checked it against my setup and told me it doesn’t work with how I use tmux. It even found the exact GitHub issue that would have to be fixed first. On launch, terminal-browser changes tmux’s key settings for the whole server and never restores them, which would silently break keybindings I rely on.

http://github.com/zenbu-labs/terminal-browser/issues/99

@abhyrama: I’ve stopped deciding about new tools from their marketing pages alone. Claude Code already knows my setup: my dotfiles, my hooks, the rules for how work gets planned and launched. So when a tool shows up, I ask one question. Claude goes through everything about the tool (its website, its GitHub repo, its issues, its code) and checks all of it against how I work. I don’t read any of it myself. This week that was Paseo, an open-source app for running Claude Code sessions from your phone, browser or desktop. Could it replace my home-grown setup for running sessions in parallel? Claude told me exactly what it could replace in my workflow today: juggling terminal tabs to keep track of parallel sessions, and reaching those sessions from my phone. It also told me what it couldn’t: the machine that runs the sessions, the planning rules every piece of work goes through, the way I set up a fresh checkout for each task, and everything I’ve built outside session management. Both lists, checked against my setup, made the decision for me. An evaluation that used to take an afternoon of reading is now fast and automated. It’s also personalized, which no feature comparison is. The question was never “is this tool good”. It was “does it fit what I already run, and make it better”.

Busy weekend managing 15 agents.

I see everyone raving about Instinct. It feels magical when an agent handles things end-to-end without you having to specify every detail.

I haven’t used Instinct, but my personal agent does most of this. The secret is context engineering: give the agent the data (Gmail, Drive, etc.), permissions, and tools (like browser use) it needs, and even a barebones agent can go a long way.

@mvanhorn: Instinct’s “can do” go through walls attitude is still the best of any personal agent. It just did my son’s Global Entry application with no inputs from me. Found every country he’s been to, passport details and did everything on it’s own. WILD experience.

In the age of AI, owning your data is a must for the tinkerer and the curious.

I have set up a script that pulls my Pixel Watch vitals out of Google into a SQLite file on my own box. The file holds resting HR, HRV, SpO2, sleep, steps, and intraday heart rate: about 470 days and 17M heart-rate rows.

Claude queries the file directly using a custom medical paper-search tool I built.

I do Norwegian 4×4 interval training on a stationary bike. I recently came back from 17 days off, and the first session felt brutal. I was curious whether the data agreed, so I asked Claude straight away.

It did. It was the hardest of my recorded rides; across all of them, a longer break meant more time at high heart rates. It also noted that my resting heart rate had drifted up during the break, providing plenty of statistics along the way. The paper search told me this pattern has a name—detraining—and tied each finding back to clinical studies.

Six days later, I rode the same bike at the same resistance and asked Claude to run the analysis again. This time I held a higher cadence at a lower heart rate: more work per heartbeat.

I could only do this data detective work because I had the raw data on my own disk, alongside an AI that could reason over it and verify the findings against medical research.

Whenever I see any video today that makes me go “wow,” the first question I have in my mind is: Is this AI-generated?

The shuffle dance in https://www.youtube.com/watch?v=_D5LPxkrfqo is “wow.” The first thought that came to my mind was: Is this AI-generated?

I replaced a paid dictation app with one I vibe coded in two days. The code was the cheap part — a model wrote most of it. What took the time was recording a corpus and measuring three speech models before writing a line of the app.

https://abhyrama.com/replacing-wispr-flow/

Stress testing my always-on Hetzner agent box: 24 agents at once on 4 vCPUs. Load average 2.09 — the CPU is bored. RAM is the wall: 6.8 of 7.7 GB gone.