digest

March 17, 2026

2026-03-17
3 itemsX×3
X

@brian_scanlan thread (24): @brian_scanlan: We've been building an internal Cla…

@brian_scanlanthreadx

TL;DR. Intercom built an internal Claude Code platform with 13 plugins, 100+ skills, hooks enforcing PR workflows, and a read-only Rails production console exposed via MCP — used heavily by non-engineers without incident.

Takeaways

  • Read-only prod Rails console via MCP: read-replica only, blocked tables, mandatory model verification, Okta auth, DynamoDB audit trail — top users turned out to be designers, CS engineers, and PMs.
  • Full OpenTelemetry instrumentation of every Claude Code lifecycle event (14 types → Honeycomb); SessionEnd hook runs Haiku over transcripts to auto-classify gaps and file pre-filled GitHub issues, closing the loop into new skills.
  • Hooks enforce process at the shell level: PreToolUse intercepts gh pr create to demand business intent, blocks edits to merged branches, and after 5 permission prompts triggers a 14-day transcript analyzer that auto-promotes safe commands to settings.json.
  • Flaky-test fixer is a 9-step forensic workflow with a 20-category taxonomy and hard rules ("NEVER skip a spec as a fix"), sweeping for sibling instances of the same antipattern across the suite.
X

Lessons from Building Claude Code: How We Use Skills

@trq212articlex

TL;DR. Skills work best when they cluster into clear categories—library refs, verification harnesses, data fetchers, or process automations—and lean on scripts and assets, not just markdown.

Takeaways

  • "Just markdown files" is the wrong mental model: the power is in folders with scripts, assets, and dynamic hooks the agent can run.
  • Verification skills (Playwright/tmux drivers that record video or assert state at each step) are worth dedicated engineering time—they keep Claude honest about whether code actually works.
  • The best skills fit cleanly into one category; the confusing ones straddle several—use the taxonomy as an audit for gaps in your org's skill library.
X

Patrick Collison tells people in their 20s to not move to San Francisco.

@InvestLikeBestthreadx

TL;DR. SF has gotten too safe, too consensus-driven, and is now minting cautious founders who'd rather wrap OpenAI than risk going broke.

Takeaways

  • SF's consensus bias is a feature when you're ahead of the curve, a bug when it disconnects you from how the rest of the world actually lives.
  • Pre-2008 founders operated "on the edge of the knife"; today's 23-year-olds get $3M seed rounds with "founder" as a resume hedge — no real downside, no real intensity.
  • Emerging-market founders build under genuine constraint, which breeds a creativity SF and NYC can't replicate from abundance.
  • "Starting companies is just too f**king safe" — the safety is attracting people who want to be employees, not founders.