1M conversations on Claude analysed

Health and wellness ranked at #1

https://www.anthropic.com/research/claude-personal-guidance

https://x.com/astnkennedy/status/2049942673608352236?s=46&t=nluSw7cwYGx8zFxfoYi5Iw

I can totally relate to this, I only multi task so much cognitively. When I am in flow state, I can go for hours before I need to totally disengage for awhile.

Personally, How do you all deal with the multi tasking fatigue? Or am I just getting old? 👨‍🦳👴

https://x.com/petergyang/status/2051129249348894754?s=46

Good thread on Openclaw vs Hermes

Ways my AI agent actually saves me time and helps to make me $$$:

Every week I get sent a bunch of market outlook videos from a curated list of people I follow. Sometimes it’s a quick few minutes, sometimes it runs up to an hour.

The reality is, I don’t always have the time to sit through all of them. And when I miss them, I fall behind.

So instead of trying to “be more disciplined,” I built a workflow where my agent:

watches out for those videos, and
auto-transcribes them (using Google Gemini paid api), then sends the key actionables straight into my Telegram group

So even if I don’t watch everything, I still stay in the loop for the markets and place some position trades.

You can apply that towards your domain.

Many takeaways from this talk, but this one stood out the most for me:

Karpathy is skeptical of the simplistic "agents will do everything" narrative.

His preferred model is partial autonomy.

He talks about keeping Al "on a leash", building systems where the Al generates and the human verifies

Don't jump straight to total delegation.

Instead:
let Al do pieces, keep human approval at key checkpoints, compress human verification cost, increase autonomy only where reliability earns it

https://youtu.be/96jN2OCOfLs?si=0Hu5-n40cJ5jUnMs

Chief Justice Sundaresh Menon says 92 percent of new lawyers were already using AI in their work.

If these foundational tasks are increasingly outsourced to machines, and if it becomes uneconomic to have them performed by young lawyers, then we must confront a serious question: how are we going to redesign our workflows and processes to ensure that our young lawyers acquire the instincts, the discipline, and the professional judgment that these very tasks once helped cultivate?"

https://www.channelnewsasia.com/singapore/2026-mass-call-high-court-chief-justice-ai-work-practices-6067356

Just shipped Prompt Architect — open source on github:

If you struggle with how to prompt, resulting in dissatisfying outputs - this is for YOU!

A Claude plugin built straight from Anthropic's "Prompting 101"
workshop framework (https://www.youtube.com/watch?v=ysPbXH0LpIE)

Skill works Interview-style: asks 4 questions about your
task, hands back a structured prompt with a test plan.

No more trial-and-error with your prompts.

Works in Claude Cowork (desktop) and Claude Code (CLI).

github.com/hosanxiv/prompt-architect

Try it out!

ChatGPT on a roll - now available as an add-on in Excel and Google Sheets.

https://chatgpt.com/apps/spreadsheets/

https://x.com/claudeai/status/2052060691893227611?s=46

Claude subscribers, Go forth and create something with the usage refresh!

https://blog.google/products-and-platforms/devices/fitbit/fitbit-air/

Google just released the Fitbit Air - a 24/7 screenless health tracker - their version of a whoop killer.

Would you cop?

Something light for the weekend

Try /radio on Claude code

https://x.com/claudedevs/status/2052818282294726699?s=46

https://x.com/livinoffwater/status/2052710806568267897?s=46

If you have a hardware device, try this out! Codex Pets is quite kawaii

I shipped something this weekend that scratched a very personal itch - Kiasumiles

my wife keeps asking me which card to use at the cashier. I keep answering confidently. i'm not always right. the problem isn't laziness — it's that Singapore credit card earn rates depend on MCC codes (Merchant Category Codes), which don't always match what the merchant looks like. a café might code as fast food. a supermarket might code as a department store. tap the wrong card and you've lost miles you'll never get back, and that could possibly cost you a trip to Bali, or even Tokyo.

so I built KiasuMiles — an MCP server that runs locally on your machine. you tell it which cards you carry (in plain English, once), and from then on you just ask your AI agent - and it’ll tell you which card to use, based on the cards you have.

"what card at Sheng Siong?" → UOB PP Visa, 4mpd, Cap: SGD 600/month, high confidence.
"best card for Grab?" → it tells you.
"which card at Sushi Tei?" → it tells you that too, with the cap and a confidence level.

no API keys. no cloud. one install command:
pip3 install kiasumiles-mcp && kiasumiles-setup

works on Claude Desktop, Claude Code, OpenClaw, Hermes.

48 cards supported across all major SG banks.

full build story — including the bugs I caught, the 150 lines I deleted, and the time the tool was confidently wrong at Sheng Siong — on my substack: https://hosanxiv.substack.com/p/every-wrong-tap-is-miles-you-earned

check it out here, do welcome your feedback: github.com/hosanxiv/kiasumiles

Would I do this project again?

honestly? probably not.

the data collection alone was brutal - cross-referencing multiple sources that all give you different numbers, none of them clearly wrong. and that's before you get into the permutations: cards with spend windows, cards where the bonus category changes per cardholder, promotional caps that expire without announcement.

got it done. wouldn't sign up for it twice. 😅

I think we may be onto something here! another person’s concurrently building something similar.

anyone want to be a co-contributor to Kiasumiles? 🤪

Really good read by YC CEO on how to make your agent better.

How he used it to ingest books and identify ideas that can be applied specifically to your life, prep for meetings.

Gonna be trying this book-mirror skill out!

Book-mirror skill extracts book chapters then maps every idea specifically to my life context, family history, YC work, and therapy, producing 30,000-word personalized brain pages in 40 minutes.

https://x.com/garrytan/status/2053127519872614419?s=46

What Meta used to do with training their algorithms on how something goes viral, by looking at how the brain reacts, is now available for the masses.

Might be slop for its initial release, but I think you can take the same idea and reiterate.

https://x.com/higgsfield/status/2053139109074657482

https://github.com/mvanhorn/printing-press-library/tree/main/library/devices/whoop

If you use whoop, install this on your claw/hermes

Sample prompts:

“Every morning, use pp-whoop and give me a one-paragraph readiness brief plus a green/yellow/red training recommendation.”

“Pull yesterday’s WHOOP data and append a summary to my daily journal.”

“Combine my WHOOP recovery with my calendar and tell me which meeting-heavy days correlate with bad sleep.”

“Compare my WHOOP strain and recovery before and after travel days.”
“If my recovery is below 40%, draft a lighter training plan for today.”

Interesting list. think at this rate we are headed - AGI might already happen 2027/2028.