Claude Code: skills and sub-agents

• The alias is the whole configuration:

import { defineConfig } from "vitest/config";

export default defineConfig({
  test: { alias: { "@/": new URL("./src/", import.meta.url).pathname } },
});

• The file is .mts because Vitest 5 warns when it loads an ESM vitest.config.ts from a package without "type": "module", and suggests .mjs or the package setting. .mts needs no change to package.json.
• When a client component needs events or state in a test, add jsdom and a DOM testing library then.
• A test sits beside its component: name.test.tsx next to name.tsx.

• The skill's instructions also name 404 for a missing resource and 409 for a conflict, for routes that have a resource; an echo route has neither.
• disable-model-invocation: true keeps a skill out of Claude's own tool choices, so /pre-deploy runs when you ask and not when Claude decides to.
• An inline command, an exclamation mark followed by the command in backticks, runs before the model reads the prompt, and its output replaces the placeholder. That is why /pre-deploy needs no tool calls.
• $ARGUMENTS is the whole argument string and $0 its first word: for /api-route echo POST, $0 is echo.
• A skill runs without a session too. The screenshot above is the output of claude -p "/pre-deploy", which printed the result and exited.

• model: sonnet on each agent: 3 agents run per review.
• tools without Edit or Write makes the reviewers read-only by construction, whatever the prompt says.
• "Report findings only" and "if nothing is wrong, say so in one sentence" keep the merged review short. The skill tells the aggregator not to add findings of its own, so every item traces to a reviewer.
• An agent's description is how the session decides when to delegate unprompted. The 3 descriptions say "use when reviewing a pull request or branch" for that reason.
• Cost: each agent reads the whole diff, so a review costs about 3 times a single pass. Keep the diff to one pull request.

• The hook entry:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.file_path' | xargs pnpm exec prettier --write --ignore-unknown --log-level warn"
          }
        ]
      }
    ]
  }
}

• PostToolUse runs after the edit, so it can fix or report, not block. A hook that must stop an action goes on PreToolUse.
• Exit code 2 from a hook sends its stderr to Claude as feedback; any other non-zero exit shows it to you only. The formatter exits 0, so Claude sees nothing.
• --ignore-unknown makes Prettier skip files it has no parser for, instead of failing the hook on a .png.
• .claude/settings.json is committed and applies to everyone; .claude/settings.local.json is personal and ignored by git.