Blog
Latest posts, updates, and notes.
When to clear the context instead of compacting it
Three studies and Claude Code's own docs show when to clear the context instead of compacting it.
Read more
Four rules that make an AI review loop converge
Four rules from the Claude Code team that stop an AI review loop re-reading the same code, and how to switch each one on…
Read more
When agents send changes faster than git can land them
GitHub and Anthropic show how agent merge rates, merge queues and CI checkouts strain git's full-copy design, starting w…
Read more
Jev: The Promise and Limits of a Model Built to Decide
TypeSafe built Jev to make decisions, not write text. We explore how it works and test the company's claims against inde…
Read more
Spend on the harness before the judge
Anthropic's red-team study of Auto Mode shows how an LLM judge's harness shapes what the judge sees.
Read more
Anthropic's Hacker-Opus: When the Checker Becomes the Task
Anthropic's Hacker-Opus experiment shows how visible graders, score-focused prompts, and agent loops can push reward-see…
Read more