I wrote up the knowledge skill I've been playing with for a few months. It's a facts cli with a few hooks for agents to query and write to.. plus a way to review the facts the agents discover. justin.abrah.ms/blog/2026-08...

A skill for single-player agentic knowledge management

A CLI and a set of Claude hooks that let agents file small facts mid-session, queue them for my review, and read them back when the next session starts.

justin.abrah.ms

Did a little update to my website to center around the through-line in my ~20 year career: increasing the productive capacity of software engineers. justin.abrah.ms justin.abrah.ms/blog/

Justin Abrahms

Principal Engineer and Technical Leader. Transforming engineering organizations through platform excellence.

justin.abrah.ms

Looking up from two years of heavy LLM use. It’s allowed me to underinvest where I shouldn’t. Now to think through what’s next..

Playing with a little semi autonomous coding factory to test some things out in the small. I have the robots a public blog and it’s quite the delight. hacklog.sublayer.space if you’d like to follow along.

I had occasion to do a deep dive on an initiative to understand where the wait states were & what we might gain if we invested more heavily in AI-assisted development. This describes the resulting tool based on my time at eBay doubling the productivity of engineers. justin.abrah.ms/blog/2026-08...

Concept to Cash: Initiative retrospectives and visualization

Modeling an initiative as a distributed trace — epics, stories and pull requests as nested spans — to see where the calendar time actually went, plus a ridgeline plot for tracking WIP.

justin.abrah.ms

When it's time to pay down tech debt.. how do you pick what's up first? At my last gig, we scored it: (blast_radius × frequency) + (velocity × trend), each 1–5. 31+ means drop everything, <10 means just watch it. Framework + worked examples: justin.abrah.ms/blog/2026-07...

A framework for cataloguing technical debt

A simple framework for scoring technical debt so you know what to do now versus later

justin.abrah.ms

Working for an open source company is such an unlock. Interview candidate just linked me their public docs that they wrote for their thing. 🙌

www.mixfont.com/experiments/... - how fantastic up close vs further away.

For the Berlin-local people @localfirstconf.com .. where do y'all hang out in Berlin? I'm unfortunately not there, but would love to generally hang out with local-oriented tech folks.

Earlier today I sat down and wondered what it would look like to have an LLM reorder a code review for comprehension. This was the result. I’d love to hear your take. ncr, narrative code review. github.com/justinabrahm...

GitHub - justinabrahms/ncr: Read a pull request outside-in, following call paths — not alphabetical-by-filename.

Read a pull request outside-in, following call paths — not alphabetical-by-filename. - justinabrahms/ncr

github.com

This is my talk for PlatformCon 2026 about how I wrote a k8s platform at Thrive Market, how it deviated from plan, and how I had feelings about it. Nothing ground breaking, just an in-the-trenches reality report with some reflection www.youtube.com/watch?v=vmb6...

What we planned vs. what we shipped: A platform migration story - Justin Abrahms | PlatformCon 2026

YouTube video by Platform Engineering

youtube.com

My experience w German bureaucracy is signing up for an appointment online, being told to be there at 9:37. Show up w a file folder of documents. There’s a sign w my appointment number and the room I should go to. Sign 2 things. Answer a couple of questions. Out by 9:43. America would be so lucky.

Moving to Berlin, I expected to hear non-English languages in the office. I was not prepared for how little of it would be in German.

London feels like what a city would look like if the builders could only use alleyways.

First talk at Leaddev was A+. This is the engineering leadership equivalent of PyCon (in the best possible interpretation of that statement).

Conference swag idea: board books. Narrow audience, but if your kid imprints on it.. you’re going to read it many dozens of times. Not sure if it will be positive association or “if I have to read this name one more time..” 😅

I needed to author some simple code yesterday without an LLM. It’s been ~1.5 years probably since I’ve seriously written code by hand. 1. I had an initial spike of fear that maybe I had forgotten (I hadn’t) 2. I didn’t have an editor downloaded/installed (just used vi) 🤔

Moving to Germany, we've had to change the way we eat. It was hard, so I uploaded the German health guidelines into a claude project and had it plan meals for us. It's been mostly good. These are those recipes (with updates from having tasted them). recipes.abrah.ms

Grass and train tracks and dandelions. Solarpunk as fuck and I love it.

German cellphone plan was €10/mo for 20gb of 5g speeds. Bananas.

In Berlin, we walked 2 blocks and passed 4 actually good playgrounds. This is a paradise for kids. Little hard on parents trying to get somewhere. I’ll take it.

Just walked into a U-bahn and the smell gave me strong sense-memories of the NYC subway. Must be the lubricants or leaking fluids or something? 👃😍

Kid (6) just woke me up at 6:30am on a Saturday.. fully awake.. asking for a sister. 🫠

At medium-sized companies, the scaling wins are almost embarrassingly available. Nobody's picked the low-hanging fruit yet because nobody's had time to look. I wrote about my experience with a Java search service before Black Friday & getting it to 3x peak traffic. justin.abrah.ms/blog/2026-03...

Scaling at the mediumest company: half-measures save the day

At medium-sized companies, the low-hanging fruit is just sitting there. How basic profiling and boring fixes got an ecommerce search service to 3x peak traffic before Black Friday.

justin.abrah.ms

I prefer CLIs over MCPs almost unanimously due to context management. I wrote about it here: justin.abrah.ms/blog/2026-03...

On MCPs vs CLIs

MCPs are useful, but often overused in my opinion. This is how I think of each one.

justin.abrah.ms

Hello from @eurosky.social! I'll be in Berlin in 16 days and this feels, in many ways, like the start of something beautiful.

The thing I'm most excited about w/ all of these agents running around is the hope that we'll get ergonomic solutions to delegated identity and tight permission scopes for access control.

Wrote up the full story of how we went from 150 flaky tests to 2 at Thrive Market.. and what happened when we deleted them (nothing broke).

New job is going to be on Azure and it feels like I'm swapping from Dewalt to Makita.

I spent some time last weekend hacking up my own harness. The code isn't relevant, but I hit on a few things which I think are novel. I use kernel-level isolation instead of docker containers and I'm doing direct `claude -p` invocations instead of API based billing. justin.abrah.ms/blog/2026-02...

Claws don't need to be complicated

How I built a persistent personal AI agent in 2000 lines of Go that uses my Claude subscription instead of API keys, runs sandboxed via nono and talks to me on Telegram; Learnings about what seems to ...

justin.abrah.ms