Steven Xu

I joined Microsoft in 2015 as an intern and started in their finance rotation program (1.5 years). Post-acquisition, I was on LinkedIn's finance team for 3 years. I've just transitioned to the #productmanagement team. This profile's for #learninginpublic!

For others, it’s just fast leverage. They can download my curated Skills as plugins, point their Claudes at my analysis and context repos, and accelerate their own work and understanding

For me, these compound my ability to do quick analysis; I point Claude to the most salient datasets and the chatbot is now a data analyst extension of myself. I’m able to answer queries for ICs and Leads that were previously unexplored curiosities due to low bandwidth.

I’ve always appreciated documentation as a way to capture past insights. Context Repos or the Andrej K “knowledge base” is now a go-to for me.

The bottleneck for good decisions at big companies usually isn't intelligence; it's bandwidth. So many hypotheses wait in queue. Questions go unanswered for ages. LLM-assisted analysis enables credible answer in hours. That speed changes what questions you even bother to ask.

I wrote SQL against 2 years of data, decomposed retention by cohort, and compared across every dimension that mattered. The "doomsday scenario" everyone feared was further than we thought. This ended up a mythbusting doc that directly reprioritized a multi-quarter strategy.

Claude unlocked analysis that was missing for weeks. Our team was debating a change that could've been existentially urgent. The fear: massive revenue loss. The hope: upside. But there was no coherent narrative. As a product manager, I did it in 4 hours with Claude.

Ultimately the slop wasn’t worth automating but it was my first use case and pretty fun. It got me excited to try new use cases. This was just 8 weeks ago.

First project was slop, but still useful. Claude read user feedback JIRA tickets, categorized them, flagged patterns that weren’t visible in individual tickets, and how often they recurred. I’m not even directly responsible for ticket triage, but it was so easy I just did it.

I keep a log of all my Claude use cases and I’m genuinely surprised at how fast the learning and application has ramped up for me so I’m going to start listing a few out from exceptionally simple slop to analytical outcomes that were previously unavailable.

I fwded to folks who actually do the triaging. They were ahead of me with an automatic reviewer PLUS PR draft to address. Pattern recognition was still useful. Timeliness to insight was an issue (I aggregated daily but tickets flowed regularly)

New post on the LinkedIn eng. blog where I have spent the majority of my time in the last year working on semantic job search. I worked w/ team on the cross-encoder small language model as well as experimenting with multiple objectives & ranking! linkedin.com/blog/engineeri…

Vibe coded: a research summarization tool to quickly assess whether a paper/PDF might be relevant to your domain. Technical novelty + business impact summaries included. Papers uploaded go to a public community feed. Lmk if you try it out! paperlens.replit.app

Not having to log onto VPN >>>>>

Product suggestion for @Uber: check the app languange of the recipient for automated messages. I sent a "Be Right There" and received a "好的。明白了！" Makes sense in UI to send in your language. Doesn't make sense to receive messages not in your language, IMO.

Not totally representative of how I spend my time, but I like this chart.

I think a huge indicator of strong partnership can be how you would feel if they left. I would strongly consider leaving if my stellar engineering manager partners decided to leave.

1/x 🧵 I was thinking about my portfolio and it occurred to me that @JoinCommonstock has a unique position to create a product feature unavailable (to my knowledge) on other platforms.

No real deep learning yet. Just a realization that permanence must surely play a role in Google's ranking, whereas it may not be as necessary in other ranking algorithms

On the other hand, job listings don't last forever like webpages, and weighting prior traffic (e.g. views, applies) might be more harmful to jobseekers because it means there's increased competition to be hired for that job.

Showerthought I had today: LinkedIn job listings are ranked differently than say, Google, because job listings are ephemeral. I would guess there is a weighting to the existing traffic a webpage gets which keeps it relevant to the top of results.