cos-review

/cos-review — Chief of Staff Review

Diagnose, build, and grow your personal AI operating system.

Your AI is your Chief of Staff. This skill is the performance review.

Detect Mode

Check what the user typed:

• /cos-review setup → run Setup Mode
• /cos-review status → run Status Mode
• /cos-review (no arguments) → run Review Mode

---

MODE 1: Setup (first run)

Run this the first time, or when rebuilding the system from scratch. It scans everything, analyzes usage, and guides the user through fixing their biggest gaps.

Step 1: System Scan

Scan automatically — don't ask, CHECK. Report what exists and what's missing.

Important: Use Read, Glob, and Grep tools (cross-platform) instead of shell commands like find or grep. Only fall back to Bash when no dedicated tool works. Detect the OS first (uname on Mac/Linux, or check for Windows paths) and adapt commands accordingly.

L1: Knows You

• Read ~/.claude/CLAUDE.md — exists? word count?
• Search for SOUL.md:
  1. Check ~/.claude/SOUL.md first (the standard global location)
  2. If not there, Glob **/SOUL.md in home (max depth 5, exclude node_modules/.git/Library)
  3. If found elsewhere but NOT in ~/.claude/: report "SOUL.md exists at [path] but is not globally accessible. Recommend: copy to ~/.claude/SOUL.md"
  4. Check if ~/.claude/CLAUDE.md references SOUL.md — if not, flag: "SOUL.md won't be auto-loaded unless CLAUDE.md references it"
• Search for user-profile.md — same 4-step approach as SOUL.md
• Count project-level CLAUDE.md files — use Glob **/CLAUDE.md from home (max depth 4, exclude node_modules/.git)
• Check memory — use Glob ~/.claude/projects/*/memory/*.md and count entries

L2: Has Sources

• MCPs and plugins are configured in THREE locations — check ALL of them:
  1. ~/.claude/.mcp.json — read and parse the mcpServers key (standalone MCP servers)
  2. ~/.claude/settings.json — read and parse the enabledPlugins object (key:boolean dict, e.g., "gmail@claude-plugins-official": true)
  3. Project-level .mcp.json files — Glob for them in the current project
• For each source found, try a lightweight test call (skip if tool not available):
  • Gmail: try gmail_get_profile or gmail_list_labels
  • Telegram: try list_chats (limit 1)
  • AnySite: try duckduckgo_search with "test"
  • Calendar: try listing today's events
• Ollama: run ollama list (works on Mac, Windows, Linux). If command not found, mark as not installed.
  • If installed, check for both local models (e.g., qwen3.5) and cloud models (e.g., kimi-k2.5:cloud, qwen3.5:cloud)
  • Note: cloud models via Ollama are fast and cheap but data goes to a third-party provider. Local models are fully private but require a strong GPU — slow on laptops.
  • If not installed and user has privacy or cost concerns, suggest: brew install ollama then ollama launch claude --model kimi-k2.5:cloud for cost savings, or local model for full privacy
• ActivityWatch (focus/productivity tracking): run curl -s http://localhost:5600/api/0/info 2>/dev/null via Bash. If responds, it's running. If not, check if aw-qt or aw-server process exists. If installed but not running, note it. If not installed, suggest as optional: "ActivityWatch tracks your screen time and focus patterns — your /morning can use it to report deep work hours and app usage. Install from https://activitywatch.net"
  • If ActivityWatch IS running, check for the analysis skill: Glob ~/.claude/skills/activitywatch*/SKILL.md
  • If AW runs but no analysis skill: "You have ActivityWatch collecting data but no way to analyze it. Install the analysis skill from https://github.com/BayramAnnakov/activitywatch-analysis-skill — it calculates focus scores, detects death loops (repetitive app switching), and generates weekly insights. After installing, run it once with defaults then review uncategorized apps to calibrate for your workflow."
• Report which sources are installed, which respond, which have expired tokens