Drew Breunig
@dbreunig.bsky.social
My not-so-monthly newsletter just went out, covering a book on context engineering I'm editing for O'Reilly, upcoming events you should checkout/attend/submit a paper to, and recent writing. buttondown.com/dbreunig/arc...
A New Book, Events, an Explainer, & More!
It's a packed issue: I'm editing a book on context engineering, there are upcoming events, writings on coding agents and spec-driven development, and more...
buttondown.com
February 15, 2026 at 6:50 PM
My not-so-monthly newsletter just went out, covering a book on context engineering I'm editing for O'Reilly, upcoming events you should checkout/attend/submit a paper to, and recent writing. buttondown.com/dbreunig/arc...
All the content created by clawdbots isn’t passing the test I keep applying to image, music, and video gen tools: if this wasn’t created by an AI would it be novel?
February 15, 2026 at 1:59 AM
All the content created by clawdbots isn’t passing the test I keep applying to image, music, and video gen tools: if this wasn’t created by an AI would it be novel?
Srihari Sriraman and I took a deep dive into coding agent system prompts to understand their structure, similarities, and differences. www.dbreunig.com/2026/02/10/s...
How System Prompts Define Agent Behavior
System prompts matter far more than most assume. A given model sets the theoretical ceiling of an agent’s performance, but the system prompt determines whether this peak is reached.
www.dbreunig.com
February 13, 2026 at 7:18 PM
Srihari Sriraman and I took a deep dive into coding agent system prompts to understand their structure, similarities, and differences. www.dbreunig.com/2026/02/10/s...
Reposted by Drew Breunig
Two Harbors Catalina 1, LosAngeles County, CA
🗺33.4413, -118.4861 🧭264° ⛰354 ft
https://ops.alertcalifornia.org/cam-console/2468
🗺33.4413, -118.4861 🧭264° ⛰354 ft
https://ops.alertcalifornia.org/cam-console/2468
February 10, 2026 at 11:08 PM
Two Harbors Catalina 1, LosAngeles County, CA
🗺33.4413, -118.4861 🧭264° ⛰354 ft
https://ops.alertcalifornia.org/cam-console/2468
🗺33.4413, -118.4861 🧭264° ⛰354 ft
https://ops.alertcalifornia.org/cam-console/2468
Why I’m excited about RLMs www.dbreunig.com/2026/02/09/t...
The Potential of RLMs
Handling Your Long Context Today & Designing Your Agent Tomorrow
www.dbreunig.com
February 10, 2026 at 12:58 AM
Why I’m excited about RLMs www.dbreunig.com/2026/02/09/t...
A napkin drawing of what you want AI to make is virtually guaranteed to be more interesting and I wish more people would try it.
My favorite artifact in the "you don't need AI to do art" discourse is, of course, screenwriter Dan O'Bannon's sketch of the facehugger to explain it to Giger.
February 7, 2026 at 8:53 PM
A napkin drawing of what you want AI to make is virtually guaranteed to be more interesting and I wish more people would try it.
RLMs are an incredible strategy for mitigating context rot. DSPy is the easiest way to use RLMs. Check out this overview: www.cmpnd.ai/blog/rlms-in...
RLMs in DSPy - cmpnd
Recursive Language Models are a new strategy for dealing with long context problems. We've implemented them in DSPy so you can quickly try them with existing or new DSPy programs.
www.cmpnd.ai
February 6, 2026 at 5:57 PM
RLMs are an incredible strategy for mitigating context rot. DSPy is the easiest way to use RLMs. Check out this overview: www.cmpnd.ai/blog/rlms-in...
A whenwords update, looking at more examples of spec driven development, including just-bash, monty, and Anthropic's c-compiler. www.dbreunig.com/2026/02/06/t...
The Rise of Spec Driven Development
Writing about AI, geo, culture, media, data, and the ways they interact.
www.dbreunig.com
February 6, 2026 at 5:41 PM
A whenwords update, looking at more examples of spec driven development, including just-bash, monty, and Anthropic's c-compiler. www.dbreunig.com/2026/02/06/t...
Reposted by Drew Breunig
using claude code at a standing desk feels like DJing
February 3, 2026 at 9:20 PM
using claude code at a standing desk feels like DJing
If you're doing applied AI research (especially system design, benchmarks, evals, efficiency, or ops) you should be submitting to the Conference on AI and Agentic Systems... caisconf.org/pages/cfp/
February 3, 2026 at 10:58 PM
If you're doing applied AI research (especially system design, benchmarks, evals, efficiency, or ops) you should be submitting to the Conference on AI and Agentic Systems... caisconf.org/pages/cfp/
I guess we’re gonna end up building a coalition around the constitution. Full circle.
February 2, 2026 at 3:38 AM
I guess we’re gonna end up building a coalition around the constitution. Full circle.
Anthropic comms is crazy.
“Here’s a 38-page essay about the existential threat of AI and how underprepared we are.”
*30 min later*
“Also you can now connect your B2B SaaS apps to Claude.”
“Here’s a 38-page essay about the existential threat of AI and how underprepared we are.”
*30 min later*
“Also you can now connect your B2B SaaS apps to Claude.”
January 26, 2026 at 7:15 PM
Anthropic comms is crazy.
“Here’s a 38-page essay about the existential threat of AI and how underprepared we are.”
*30 min later*
“Also you can now connect your B2B SaaS apps to Claude.”
“Here’s a 38-page essay about the existential threat of AI and how underprepared we are.”
*30 min later*
“Also you can now connect your B2B SaaS apps to Claude.”
For people wondering how they can price so high (and why many advertisers will pay for it) read this: www.dbreunig.com/2022/06/06/w...
They’re in the “shiny object phase”.
They’re in the “shiny object phase”.
January 26, 2026 at 3:36 PM
For people wondering how they can price so high (and why many advertisers will pay for it) read this: www.dbreunig.com/2022/06/06/w...
They’re in the “shiny object phase”.
They’re in the “shiny object phase”.
Probably worth revisiting this bit I wrote in 2016… www.dbreunig.com/2016/08/04/a...
January 24, 2026 at 1:59 AM
Probably worth revisiting this bit I wrote in 2016… www.dbreunig.com/2016/08/04/a...
I wanted to download a file from a Claude Code Web instance, not commit it to Github.
I don't know if you've tried this, but the first working solution it arrived at was outputting a gz file as base64 then asking me to paste it into a txt file.
I don't know if you've tried this, but the first working solution it arrived at was outputting a gz file as base64 then asking me to paste it into a txt file.
January 23, 2026 at 3:56 AM
I wanted to download a file from a Claude Code Web instance, not commit it to Github.
I don't know if you've tried this, but the first working solution it arrived at was outputting a gz file as base64 then asking me to paste it into a txt file.
I don't know if you've tried this, but the first working solution it arrived at was outputting a gz file as base64 then asking me to paste it into a txt file.
Regarding the Claude “Soul Document”: I wonder if training models on documents like these influences our perception of these models as sentient.
In many ways, LLMs are playback machines, which we talk to as human. It reminds me of facilitated communication. en.wikipedia.org/wiki/Facilit...
In many ways, LLMs are playback machines, which we talk to as human. It reminds me of facilitated communication. en.wikipedia.org/wiki/Facilit...
Facilitated communication - Wikipedia
en.wikipedia.org
January 22, 2026 at 3:16 AM
Regarding the Claude “Soul Document”: I wonder if training models on documents like these influences our perception of these models as sentient.
In many ways, LLMs are playback machines, which we talk to as human. It reminds me of facilitated communication. en.wikipedia.org/wiki/Facilit...
In many ways, LLMs are playback machines, which we talk to as human. It reminds me of facilitated communication. en.wikipedia.org/wiki/Facilit...
Dara visualization in the wild.
January 21, 2026 at 5:45 PM
Dara visualization in the wild.
An irony of the AI age is that relationships matter more. www.economist.com/business/202...
Job applicants are winning the AI arms race against recruiters
Companies have been hit by a wave of chatbot-generated applications
www.economist.com
January 17, 2026 at 5:55 PM
An irony of the AI age is that relationships matter more. www.economist.com/business/202...
Got nerd-sniped into shepherding Claude while it implemented just-bash in Python.
It's a good continuation from whenwords. The test coverage is high (the orig repo and Bash itself).
I really like this idea of emulation as an option prior to using a full sandbox. github.com/dbreunig/jus...
It's a good continuation from whenwords. The test coverage is high (the orig repo and Bash itself).
I really like this idea of emulation as an option prior to using a full sandbox. github.com/dbreunig/jus...
GitHub - dbreunig/just-bash-py: Python Bash emulation for agents, a port of vercel-labs/just-bash
Python Bash emulation for agents, a port of vercel-labs/just-bash - dbreunig/just-bash-py
github.com
January 17, 2026 at 5:46 PM
Got nerd-sniped into shepherding Claude while it implemented just-bash in Python.
It's a good continuation from whenwords. The test coverage is high (the orig repo and Bash itself).
I really like this idea of emulation as an option prior to using a full sandbox. github.com/dbreunig/jus...
It's a good continuation from whenwords. The test coverage is high (the orig repo and Bash itself).
I really like this idea of emulation as an option prior to using a full sandbox. github.com/dbreunig/jus...
AI writing feedback always feels a little off to me… Like it's been post-trained to death on linear reasoning sequences, or something.
January 17, 2026 at 12:02 AM
AI writing feedback always feels a little off to me… Like it's been post-trained to death on linear reasoning sequences, or something.
You can bet $10k on the Dodgers making the playoffs, to win $100.
Or you can put $10k in a 1 year CD at 4% and get $400.
Or you can put $10k in a 1 year CD at 4% and get $400.
January 16, 2026 at 5:49 PM
You can bet $10k on the Dodgers making the playoffs, to win $100.
Or you can put $10k in a 1 year CD at 4% and get $400.
Or you can put $10k in a 1 year CD at 4% and get $400.
It is amazing how poorly labeled and undersold that chart is given the importance of what it conveys.
The gap between the two lines is the value of a human in the loop. www.anthropic.com/research/ant...
The gap between the two lines is the value of a human in the loop. www.anthropic.com/research/ant...
January 16, 2026 at 4:32 PM
It is amazing how poorly labeled and undersold that chart is given the importance of what it conveys.
The gap between the two lines is the value of a human in the loop. www.anthropic.com/research/ant...
The gap between the two lines is the value of a human in the loop. www.anthropic.com/research/ant...
Terence Tao's takes on how AI helps mathematicians resonate the most with my feeling about coding agents as well.
January 14, 2026 at 5:51 PM
Terence Tao's takes on how AI helps mathematicians resonate the most with my feeling about coding agents as well.