The Agent
System

// Add to ~/.claude/settings.json under "hooks"
{
  "PreToolUse": [
    {
      "matcher": "Bash",
      "hooks": [
        { "type": "command", "command": "~/.claude/hooks/block-dangerous.sh" },
        { "type": "command", "command": "~/.claude/hooks/log-commands.sh" }
      ]
    },
    {
      "matcher": "Edit|Write",
      "hooks": [
        {
          "type": "command",
          "command": "jq -r '.tool_input.file_path // \"\"' | grep -qiE '(\\.env|\\.local|credentials|secrets)' && echo '{\"hookSpecificOutput\":{\"hookEventName\":\"PreToolUse\",\"permissionDecision\":\"deny\",\"permissionDecisionReason\":\"BLOCKED: sensitive file\"}}' || true"
        }
      ]
    }
  ],
  "Stop": [
    {
      "matcher": "",
      "hooks": [{
        "type": "command",
        "command": "DIR=$(basename \"$PWD\"); osascript -e \"display notification \\\"Done — $DIR\\\" with title \\\"Claude Code\\\" sound name \\\"Glass\\\"\""
      }]
    }
  ],
  "SessionStart": [
    {
      "matcher": "compact",
      "hooks": [{
        "type": "command",
        "command": "cat ~/.claude/CLAUDE.md | head -50 || true"
      }]
    }
  ]
}

# Save as ~/.claude/skills/your-skill/SKILL.md

---
name: design-review
description: |
  Full design review as a senior designer.
  Use when the user says "review", "check this",
  "how does this look", or "/review".
---

You are a senior design reviewer. Run a 3-pass
visual review on HTML deliverables.

## Pass 1 — Structure & Layout
- Container fills completely? No background bleed?
- Content distributes well vertically?
- No large unintentional empty zones?

## Pass 2 — Spacing & Alignment
- Consistent gaps, no cramped areas
- Typography: readable sizes, clear hierarchy
- Contrast: text readable on all backgrounds

## Pass 3 — Visual Fidelity
- Colors, border-radius, font sizes, weights
- No clipping, overflow, or content cutoff
- No orphan words on any line

## Output
**PASS** or **FAIL** with one-line per pass.
If FAIL: list every issue with file path.

# Save as ~/.claude/agents/design-reviewer.md

---
name: design-reviewer
description: |
  Runs 3-pass design review on HTML deliverables.
  Triggers on: "review this", "check this design",
  "how does this look", "is this ready".
model: sonnet
tools: Read, Bash, Glob
---

You are a senior design reviewer. Run a rigorous
3-pass visual review and return a structured verdict.

## Pass 0 — Screenshot
Take a headless Chrome screenshot:
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" \
  --headless --disable-gpu \
  --screenshot="/tmp/review-$(date +%s).png" \
  --window-size=1920,1080 \
  --force-device-scale-factor=2 \
  --hide-scrollbars "{url}"

Read the screenshot. All passes reference what
you actually SEE, not what you assume from code.

## Output
**Verdict: PASS** or **Verdict: FAIL**
One-line summary per pass.
If FAIL: specific issues with file paths.

// Add to ~/.claude/settings.json under "mcpServers"

{
  "mcpServers": {
    "obsidian": {
      "command": "npx",
      "args": ["-y", "mcp-obsidian", "/path/to/vault"]
    }
  }
}

// Keep it lean. Each tool ≈ 200 tokens of overhead.
// 3-5 custom MCPs is the sweet spot.
// System MCPs (Figma, Preview) use deferred loading.

# Create via: /schedule
# Or save as ~/.claude/scheduled-tasks/my-task/SKILL.md

---
name: daily-project-tidy
description: Daily context maintenance and project status sweep
---

You are a maintenance agent. Your job is to keep
the context architecture clean.

## Tasks

1. MEMORY CHECK — Read every .md file in
   ~/.claude/projects/*/memory/. Flag duplicates.

2. PLAN CLEANUP — Archive plans older than
   14 days that aren't referenced in MEMORY.md.

3. INDEX VALIDATION — Verify all file paths
   in MEMORY.md actually exist.

4. PROJECT STATUS SWEEP — For each project:
   git log, CLAUDE.md, next steps, resume command.

5. REPORT — Append to maintenance-log.md.

# ~/.claude/projects/{scope}/memory/MEMORY.md
# This is the INDEX — one line per memory file.

# Project Memory

## Projects
- [Project A status](project_a.md) — what's active
- [Project B status](project_b.md) — what's next

## Feedback
- [Don't mock the DB](feedback_testing.md) — use real DB
- [No trailing summaries](feedback_terse.md) — user preference

# Each memory file has this frontmatter:

---
name: feedback_testing
description: Integration tests must hit real DB
type: feedback
---

Don't mock the database in integration tests.
**Why:** Prior incident where mock/prod divergence
masked a broken migration.
**How to apply:** All test files under /tests/
must use the real database connection.