Newsletter #33
Think Harder: How I Prompt #
— Thorsten Ball, Laracon US 2026
tl;dr: Thorsten Ball dissects how he prompts coding agents while building Amp — 99% AI-written. No MCP servers, no frameworks, no magic keywords: just relentlessly asking “how can the model know what I mean?” and pointing it at the exact files, screenshots, and docs where the information lives. His best trick is moving knowledge out of prompts and into the codebase via agents.md files, so prompts get shorter as the repo gets smarter.
AIPrompting CodingAgents Talk Amp #
Malleable software: Restoring user agency in a world of locked-in apps #
— Ink & Switch
tl;dr: Ink & Switch argues software has lost the malleability early personal computing promised — users became consumers of fixed apps instead of shapers of their tools. The essay lays out a research agenda for tools that ordinary people can modify incrementally, with AI as a potential unlock rather than a replacement. A foundational read for anyone thinking about where end-user software should go next.
MalleableSoftware EndUserProgramming Essay Research #
Geoffrey Litt: Dynamic Documents as Personal Software #
— YouTube
tl;dr: Litt traces his eight-year rabbit hole from live programming (margin notes that show what your Ruby code returned last run) through Ink & Switch’s Embark to Notion, where dynamic documents are becoming real product. His core idea: start from the notes people already write, then make them gradually more alive — browsable, then authorable, then programmable, like HyperCard’s on-ramp. Ends with a practical hook: an “explain diff” skill that turns agent changes into interactive documents with a spaced-repetition quiz you must pass before merging.
PersonalSoftware DynamicDocuments LiveProgramming Talk #
Immediate-Mode Graphical User Interfaces #
— Casey Muratori, 2005
tl;dr: The 2005 lecture that named immediate-mode GUIs: instead of retaining widget objects and syncing state, draw and interrogate the UI every frame directly from your data. Casey shows how this eliminates an entire class of state-sync bugs and makes interfaces radically simpler to build. Twenty years on, the idea powers Dear ImGui, Rust’s egui, and most dev-tool UIs — this is the origin document.
GUI ImmediateMode SoftwareHistory Talk #
Getting Better at Working With You: Compiling User Corrections into Agent Memory #
— arXiv
tl;dr: This paper treats every user correction of an AI agent as a compilation target: feedback gets distilled into durable, reusable memory rules instead of living only in chat history. The result is agents that stop repeating the same mistakes across sessions. Directly relevant if you’re building agent memory on top of your own correction patterns.
AIAgents Memory Feedback Paper #
What’s a durable skill software engineers need to learn in 2026? #
— YouTube
tl;dr: Agent workflows change constantly — tools, models, and frameworks rotate every few months — so the durable skill isn’t any specific stack. It’s the judgment to decompose problems, verify agent output, and steer systems you didn’t write line-by-line. A useful framing for where to invest learning time when everything underneath keeps shifting.
Career SoftwareEngineering AIAgents Skills #
Pydantic Monty: A Minimal Python Sandbox for AI Agents #
— Pydantic
tl;dr: Monty is a tiny, secure interpreter that runs untrusted LLM-generated Python safely — no filesystem, no network, deterministic execution. Pydantic built it so agents can compute and transform data without handing them a full runtime. A neat middle ground between “let the model exec code” and “don’t let it exec anything.”
Python Sandbox AIAgents Security #
Unmasking TeamPCP: Software Supply Chain Attacks #
— Flare
tl;dr: Flare’s threat team dissects the TeamPCP campaign, which compromised developer ecosystems to distribute malware through trusted software supply chains. The writeup covers the infection chain, the infrastructure, and the hard lesson: your dependencies’ trust is your attack surface. Worth a skim before you next npm install something unvetted.
Security SupplyChain Malware ThreatIntel #
Notable Links #
- zvec-grep — Local-first search across your workspace, built for humans and AI agents (3k stars).
- open-seo — Open source alternative to Semrush and Ahrefs (17.5k stars).
- pi-lens — Real-time code feedback for the pi coding agent: LSP, linters, formatters, structural analysis.
- pr-lens — Review code faster: draws every PR as animated architecture and data-flow diagrams.
- collie — PWA to manage herdr on the go, tailnet-accessible with push notifications.
- kache — Zero-copy, content-addressed Rust build cache for Rust and C/C++.
- lemmalog — A Datalog engine for LLM agent memory: stratified rules with provenance-tracked facts.
- Tailcat — Tailscale’s open-source CLI that connects clients and servers over WireGuard without a coordination server.
- Lucid Auth — A Rust implementation of passkey-first authentication.
- Opinions — Jasper Croome: to build software is to have an opinion.