← Latest briefing

AI Intelligence Briefing — Monday, March 23, 2026

2 top stories 28 items scanned
tools 5research 20learning 1patterns 2

Top Stories

Experimenting with Starlette 1.0 with Claude skills

Source: Simon Willison (Tier 1) | Category: patterns | Relevance: 8/10

Simon Willison documents how he’s using Claude’s “skills” feature to teach it Starlette 1.0 patterns, effectively creating reusable context that improves AI-assisted coding.

Why this matters: When you’re working with a framework that’s new or has changed significantly, AI assistants can hallucinate outdated patterns. Simon is showing a practical way to feed Claude the right knowledge so it writes correct, up-to-date code — which saves hours of debugging.

So What: This is a direct demonstration of a workflow pattern you should adopt: creating Claude skills (structured reference documents) for any framework or library where the AI tends to get things wrong. If you’re using Astro 5.x or a newer Vercel SDK, building a skill file with current API patterns will dramatically improve Claude Code’s output quality. Watch for the companion ‘Starlette 1.0 skill’ post for the actual skill document format.

Read more →

Starlette 1.0 skill

Source: Simon Willison (Tier 1) | Category: patterns | Relevance: 7/10

The actual Claude skill document Simon created for Starlette 1.0, serving as a template for how to structure framework-specific AI coding context.

Why this matters: This is a concrete, copy-able example of how to format knowledge so an AI assistant actually uses it well — think of it as a cheat sheet you give to your AI coding partner so it stops making mistakes.

So What: Use this as a template to create your own skills for Astro, Vercel’s API routes, or any library where Claude Code produces outdated code. The structure and level of detail Simon uses is a proven pattern worth replicating in your own CLAUDE.md or project context files.

Read more →

Also Notable

  • A Visual Guide to Attention Variants in Modern LLMs (Ahead of AI (Sebastian Raschka) (Tier 2)) — Sebastian Raschka provides a comprehensive visual walkthrough of attention mechanisms from MHA to MLA, sparse attention, and hybrid architectures used in modern LLMs. Understanding how the models you use every day actually work under the hood helps you make smarter decisions about which model to pick for a task and why some models are faster or cheaper than others. →
  • JavaScript Sandboxing Research (Simon Willison (Tier 1)) — Simon Willison explores JavaScript sandboxing approaches, relevant to safely executing AI-generated code in browser or server environments. If you ever let AI generate and run code — like in a tool-use or agentic workflow — you need to make sure that code can’t do anything dangerous. This research maps out how to create safe boundaries for running untrusted JavaScript. →
  • AI Agents Can Already Autonomously Perform Experimental High Energy Physics (arXiv cs.AI (Tier 3)) — Researchers demonstrate AI agents autonomously conducting high-energy physics experiments, showcasing agentic AI capabilities in complex scientific domains. It’s a striking example of how far autonomous AI agents have come — if they can handle particle physics experiments end-to-end, the same agentic patterns could handle complex business workflows too. →
  • Measuring Faithfulness Depends on How You Measure: Classifier Sensitivity in LLM Chain-of-Thought Evaluation (arXiv cs.AI (Tier 3)) — Paper shows that how you evaluate whether an LLM’s chain-of-thought reasoning is faithful to its actual decision process varies wildly depending on the evaluation method used. If you rely on an AI’s step-by-step reasoning to explain its decisions to clients or users, this is a reminder that the reasoning shown might not reflect what the model actually ‘did’ internally — so don’t over-trust chain-of-thought explanations. →
  • The Y-Combinator for LLMs: Solving Long-Context Rot with λ-Calculus (arXiv cs.AI (Tier 3)) — Proposes a lambda-calculus-inspired approach to address performance degradation in LLMs when processing very long contexts. If you’ve ever noticed that AI assistants get worse at following instructions or remembering details in really long conversations, this paper tries to fix that problem using ideas from computer science theory. It’s early-stage research but tackles a real pain point. →
  • PCGamer Article Performance Audit (Simon Willison (Tier 1)) — Simon shares a web performance audit of a PCGamer article, likely generated or assisted by AI tooling. A nice example of using AI to quickly audit website performance, but not directly related to building AI workflows or development tooling. →

📚 5 new items added to your learning queue →

Signal Scan

  • Items scanned: 28
  • Sources checked: 3
  • High relevance (7+): 2
  • Generated: 2026-03-23T11:50:26.301Z