Eric Thompson, Ph.D.

@tangming2005 @threadreaderapp unroll

If you're using Claude Code and haven't built a custom skill, start here. x.com/trq212/status/…

I've been building Claude Code skills for a month. Turns out I was missing half the picture. Thariq from Anthropic just published lessons from hundreds of skills they use internally. 9 categories, dozens of patterns.

@trq212 @threadreaderapp unroll

The Bash tool is one of the most powerful general purpose tools you can give an agent, but you also want to add in guardrails to make it safe. With the Claude Agent SDK we built a bash parser and permission system to make this easier: docs.claude.com/en/api/agent-s…

Why even non-coding agents need bash I've done dozens of calls with companies making general agents over the past few weeks and my advice generally boils down to: "use the bash tool more" Here's a concrete example from my email agent:

@tangming2005 @threadreaderapp unroll

Agents can still pick up new capabilities, but inside boundaries you set. nvidia.com/en-us/ai/nemoc…

Your AI agent can execute code and call external APIs. How many security guardrails does it have? For most people: zero. NVIDIA just open-sourced NemoClaw.

@Voxyz_ai @threadreaderapp unroll

spent an evening running the full gstack sprint on a real feature. /office-hours → /plan-ceo-review → /plan-eng-review → build → /review → /ship. three things stood out: /design-consultation analyzed my existing site, found inconsistencies, proposed updates to my design language, and built a live preview page so i could see everything before committing. after that it created a DESIGN.md file, every new component automatically followed the same system. no more "why does this button look different on every page." i normally use codex for engineering work because it's rock solid for code. but design and planning is where claude shines, so i ran the plan and build phase in claude code this time. halfway through, claude detected that codex was running on the same machine and asked me if i wanted codex to review the code before it continued. i didn't configure this. it just noticed and offered to hand off. two different AIs figured out how to collaborate without me setting it up. going from "what should i build" to "shipped PR" felt like managing a team that actually talks to each other.

@trq212 @threadreaderapp unroll

the agent SDK is the easiest way to build agents claude.com/blog/building-…

I put a lot of heart into my technical writing, I hope it's useful to you all. 📌 Here's a pinned thread of everything I've written. (much of this will be posted on the Claude blog soon as well)

@baym Remember, this was research on poisonous lizard spit!

@baym Ask your favorite LLM how many R01s contributed to GLP-1 development. The ROI for the US treasury is staggering. To get there likely took over 1000 grants based on what my LLM told me. Didn't download data sets to make it true evidence based, but I think directionally correct.

This argument, that what’s ultimately right and better may not appear to be the best solution now, is the best rebuttal I’ve heard to the current fervor for translational and applied research over basic science

@EricTopol @Waymo Wonder if there comparison data is limited to their service areas?

Latest @Waymo data based on 170 million miles waymo.com/safety/impact/

Tuesday, March 24, at 1 p.m. Eastern Webinar on the Protecting Access to Medicare Act (PAMA). Learn what changes have been implemented since the last reporting period, how they affect your lab and what you need to report by May 1. REGISTER: register.gotowebinar.com/register/67202…

@_atanas_ Ask for favorite LLM to map out the R01 grants that lead to the GLP-1s. It is a fascinating ride, and the ROI for the United States Treasury is massive. It was a convoluted road. Great to see these examples. Thanks for sharing!

Not all impactful science is obvious at the start. Some begins as risky, collaborative experiments. This is one of those stories 👇 citationclassics.com/stories/from-c… #INPST

@NotoriousS1Q3T3 Congratulations sir! Your hard work has earned you hard work you'll love.

Extremely blessed to announce that I have matched in Emergency Medicine at Stony Brook University Teaching Hospital in Long Island, New York!!!!! I am honored to join such a stellar program committed to training some of the best EM physicians! I already spoke to this some in my post from Monday, but words can’t describe how elated I am. Not matching last year broke me, and I spent most of the past year depressed after being ripped from my support system and thrust to a place where I know nobody while doing a TY that, while great, is not what I had hoped to achieve during the SOAP. And overall, I spent a lot of time questioning my place in Emergency Medicine in general. One year later, I finally know what it feels like to match to a program that I truly wanted to be at, but more importantly, that truly wanted ME. I am a more confident physician, and more confident in myself and my ability to do hard things. It took a while to see the silver lining, but I truly believe things happen for a reason—and I now know I was meant to take this path, despite it not being what I initially had envisioned. I have been blessed to be surrounded by incredible colleagues and new friends both within and outside of work, finding a small, loving, supportive community here in Fayetteville during the hardest year of my life and slowly recovered from the trauma of the 2025 Match and SOAP process. And, I have had such incredible support from new and existing mentors as I worked tirelessly to achieve this goal. It still doesn’t feel real. I keep rechecking my email and the NRMP website to make sure there wasn’t a mistake, even though my PD has emailed us all and personally called me already. I am so incredibly proud of myself and the hard work I put in to claw my way back from failure, of my perseverance borne of my decades of martial arts training. Now, the real work begins as I look towards this exciting next chapter of becoming a BAFERD. I couldn’t be more thrilled, but I also couldn’t be more thankful for the challenges that have led me to this point. Time to begin planning my move…but first, time for some well-earned celebration and PTO the next few days. LET’S FREAKING GOOOOO!!!!!! 😎🏥🚑🚨🩺🩻👨🏼‍⚕️ #Match2026 #EMbound #BAFERD

@EricTopol @NatureMedicine 6/6Guideline-mandated for 8 years. Still not reaching half of community patients. Reimbursement denial rates rising. Payment cuts coming. The evidence for NGS is not the barrier — the system is.

@EricTopol @NatureMedicine 5/6It's about to get harder. PAMA cuts of up to 15%/yr resume in 2027. The moratorium runs through Dec 2026 only. No enacted legislation fixes the rate-setting methodology. Labs absorbing repeated cuts can't sustain complex testing infrastructure. cap.org/advocacy/labor…

Why should tumor whole genome sequencing (WGS) be done for cancer? In real practice of medicine study of 888 patients with solid cancers, WGS directly led to clinical consequences in over 40% @NatureMedicine nature.com/articles/s4159…

@thenasheffect @JAMAOnc Some data from mdpi.com/1718-7729/24/5… to put a little finer point on this.

Excited to share this viewpoint @JAMAOnc on how NGS testing barriers impact timely lung cancer management! jamanetwork.com/journals/jamao…

@tangming2005 @threadreaderapp unroll - thanks for these summaries! Very helpful to have them in one place.

"Non-parametric" tests? They're just parametric tests on rank-transformed data. Not magic. This one page by Jonas Lindeloev maps it all out. I wish I had this in grad school instead of memorizing which test to pick from a flowchart. lindeloev.github.io/tests-as-linea…

Most statistical tests you learned separately are the same thing. t-test, ANOVA, Mann-Whitney, Chi-square, Wilcoxon... all just special cases of linear models. y = b0 + b1*x covers almost everything.

Moderate consumption of caffeinated coffee or tea was linked to reduced #dementia risk and modest improvements in #cognitive outcomes; no benefit was seen for decaffeinated coffee in an observational study of US adults. ja.ma/40ENw5I