Shawn Simister

I've been thinking about why verifying AI agent output feels so much harder than writing the spec that produced it. That question led me to rethink where my attention actually belongs in the process, and eventually to build atelier.dev narphorium.com/blog/decisio...

You Can CVE Everything

You Can See Everything is really good. I'm still trying to make sense of it

Great illustration of what effective voice prompting looks like. It’s a brain dump; not a spec

Watched Digger last night and really liked it. Reminds me of Asteroid City in a strange way

ParallelPilot: Supporting Coordination and Monitoring in Parallel AI Coding arxiv.org/abs/2609.33113

Looks like Gemini Argon takes more tokens to be able to deliver the same level of intelligence as Opus 5.5 (high; not max) so it would be slower and more expensive than Opus if not for the temporary 50% discount artificialanalysis.ai/models/gemin...

"I would have written a shorter report, but I didn't have enough tokens"

"Software engineering has changed radically. It is now mainly about reading code. But reading code does not scale. We need to delegate to AI." I am glad you now recognize code reading as a problem. However, the broader problem statement is not quite correct and neither is the conclusion.

TIL if you export CLAUDE_CODE_OAUTH_TOKEN it also disables Chrome MCP so Claude doesn't know how to fix simple CSS bugs anymore. Fun times code.claude.com/docs/en/chrome

Been tinkering with annotation "fusion": annotate a paper with ink; take a photo; markup is segmented, turned into structured data, and re-super-imposed on the clean original PDF.

We reimagined how computational notebooks could work with agents as the primary interface. Our new agent-first notebooks, called chats, are radically different from our human-first notebooks and yet seamlessly interoperate with them.

With Codex/Claude the bottleneck in software development is now expressing your intent. Writing the code is no longer the rate limiting step. This completely upends software development. Take issue trackers. The key problem they were solving was allocation of scarce coding resources to features.

This week's #PaperILike is "People construct simplified mental representations to plan" (Ho et al., Nature 2022). People generate ("construe") simplified task representations on-the-fly. (Robots should too!) Also: awesome first paragraph. PDF: arxiv.org/abs/2105.06948

After The Rain - Overmono

really enjoying this setup for syntopic reading with Fable

"the phoenix must be reborn with its empty spaces intact — intent includes the intent to remain open, and “immortality of intent” is only half-achieved by a system that can’t carry a dodge through the fire." wtf you talking about fable? 😂

What's new isn't the loop. It's that the controller is now expressive enough to handle unstructured goals: natural language specs, ambiguous tasks, open-ended iteration. That's the actual innovation. Not the feedback loop itself.

How to create reading experiences that "go beyond information transmission and toward reader transformation.” Great paper by @blue-phia.bsky.social @lepidopterane.bsky.social @yijunliu.bsky.social Sarah Sterman @sh1m.bsky.social @maxkreminski.bsky.social arxiv.org/abs/2606.04308 >

finally a dedicated device for the 20X per day you need to open MS authenticator 😄

What are users thinking during their interactions with LLMs? Introducing ThoughtTrace — the first dataset capturing what users think during real-world human-AI conversations. These thoughts improve user behavior prediction and model alignment, opening a new paradigm of user-centric LLM research.

Most engineers stopped reading docs ages ago. But now, agents are reading them. What if you had a way to measure freshness on every PR? Is it as easy as implementing three signals, plus a Claude Code layer? Let's find out together!

One of the most famous power-user tools in the world is switching to an AI chat interface 😬 "This will be the new Terminal. This will be the primary way most interactions are happening..." www.wired.com/story/the-bl...

For the past few years, humans have been doing “prompt engineering” to coax the best performance out of different LLMs. In this work, we explored what happens if we train an AI to do that job instead. Link to our #ICLR2026 paper: arxiv.org/abs/2512.04388 Thread:

A graph you can act on, not just stare at 🌴 @obsidian.md

Still great advice 43 years later