• Writing a good CLAUDE.md: A few good tips here around what not to put in claude.md
    • Claude can follow 150 - 200 instructions reasonably well
    • The root claude.md can contain pointers to other docs that are read on demand. (/agent_docs) Include an instruction telling claude.ai to read any it thinks are relevant
    • Don’t use claude.ai for linting. There are lots of tools that do that already cheaply and fast …
    • Claude.ai will already read local claude.md files so that’s something to keep in mind as well (But we’re living in a monorepo where claude.md files may apply to multiple project subfolders)
  • Graviton5 in Production at Honeycomb: Per-service Results From the m8g to m9g Migration: Yet another generation of amazon silicon begets honeycomb 20% performance boost for cpu bound workloads.
  • Queues Don’t Fix Overload: Yep. Back pressure and load shedding are architecture decisions that acknowledge there are limits in the system that matter. It’s nice to make them intentionally when designing a system at which point you get to decide the behaviour when work starts to pile up.
  • The radical network redesign that led AWS to forge a more resilient cloud: Interesting article about how aws went from a hierarchical datacenter network routing architecture to a random path selection one and saw much better efficiency during peak traffic times. The analysis that went into deciding this was a good thing to do was significant.
  • Production Telemetry Is the Spec That Survived: Asks the question about how to incorporate observability signals into our understanding of how a system currently works. The way we’re currently thinking about context management for an agent is probably a better fit for greenfield projects. (She talks about blackfield projects which is new to me - a project in use, under load, and potentially being replaced. With blackfield projects a testsuite starts to include elements of “this is what the system currently does” in addition to “this is what is should correctly do” which is subtly different.)
  • The Agent Harness: Turning AI Slop Into Shipping Software: How an existing codebase is made ai-navigable through test automation, and making good patterns visible, and through the use of AGENTS.md files placed intentionally.