Everything from “What Actually Happens When You Give a Dev Team Claude Code for a Day” — the open-source tools, the papers, the write-ups. Build the one ridiculous tool. Give the keepers a wall.
The PR walkthrough with the wall: a guided reading order, architecture diagrams, the why — and you can’t submit a review without opening every file.
Generate 15 full design variants, pick by eye. It’s easier to choose than to describe.
Semantic + keyword search over your own Claude Code sessions — the tool that read its own logs and rebuilt itself.
The library I made without noticing: embeddings, search, novelty, clustering for personal data projects.
The prompt → skill → tool pattern, in long form.
The project the 15 Renaissance themes were generated for.
The daily newspaper of the world, in 12 languages. Founded for one reader.
Fregnan et al., ESEC/FSE 2022 — 64% lower odds of catching a defect in the last file shown vs. the first.
Once you find one problem, you stop looking as hard. Old, human, and very much alive in code review.
Shao et al. — narrative framing measurably improves comprehension of insights.
Lee et al., Microsoft — knowledge workers using GenAI self-report less critical-thinking effort; confidence in the AI drives the drop.
Plus his Showboat tool: make the agent write up how it verified its own work.
The bill that comes due for code you didn’t read.