Skip to main content
Every client repo ships the stage-health-check skill in .claude/skills/. It runs a deterministic audit and writes a report; an agent (or you) reads the report, checks the judgement calls against the code, and ranks the fixes.

Run it

  • --skip lighthouse for a fast loop; --skip runtime,lighthouse for a static-only pass with no URL.
  • --presets mobile halves the Lighthouse time.
  • Playwright and Lighthouse install on first run into a temp tools directory, never into the site’s package.json. It drives the system Google Chrome.
  • Use the live URL. Local runs prove wiring, not performance: ISR has no local preview, and TTFB and Lighthouse numbers only mean something on the served host.

What it scores

Levels: 🔴 broken or leaking, 🟠 below the bar, 🔵 review. A dimension fails on any 🔴.

The standard lives in checks.md

The skill’s checks.md is the canonical list: every check with its why, its fix, and the Stage version that introduced the practice. A site pinned below that version is told to bump first, not that it’s broken. The file is updated whenever the platform moves — a new practice lands there in the same PR as the feature, and a fix applied on a second site becomes a check for the third. The fleet-wide view (which sites are behind, which don’t answer) is the Fleet tool; this is the per-site deep pass.