Master-level techniques and advanced workflows

Tips

Advanced techniques for getting the most out of Claude Code and the Kun engine.

Feature of the Week

/loop can babysit a PR till it merges.

Hand /loop a PR and a goal and it polls on a recurring interval — re-checking CI, fixing red checks, re-requesting review — then stops itself the moment the PR merges. No more refreshing the GitHub tab.

/loop 10m watch PR #142 — if CI is red, pull the branch and fix it; stop when merged

Omit the interval to let Claude self-pace between checks. Pair it with --dangerously-skip-permissions for a fully hands-off merge babysitter.

Keywords Are Everything

One word triggers an entire workflow. Master the keywords and you 10x your speed:

Instead of...Say...What Happens
"Start the development server"devKill port 3000 → pnpm dev → Chrome opens
"Commit and push to remote"pushLint → fix → git add → commit → push
"Deploy to production"shipBuild → verify → Vercel deploy → report URL
"Create a data table component"table usersBlock agent → DataTable → Prisma schema
"Set up authentication"authAuthJS agent → NextAuth setup

Compound Keywords

Chain keywords for complex operations:

"saas billing"     → Schema + Actions + UI + Stripe integration
"test login"       → Vitest + Playwright tests for auth flow
"handover auth"    → 5-pass QA on auth block
"motion hero"      → Framer Motion animations on hero section

Reference Other Projects

"auth like hogwarts"      → Copy hogwarts multi-tenant auth pattern
"cart from souq"          → Clone souq's cart implementation
"booking like mkan"       → Reference mkan's booking system
"table from codebase"     → Get DataTable from pattern library

Visual Debugging

Use the see keyword for visual inspection:

see                       → Screenshot localhost:3000
see /en/dashboard         → Screenshot specific page
see headed                → Open visible browser window
debug /en/settings        → Screenshot + console errors + network
check /en/dashboard       → Responsive QA at 3 breakpoints + RTL

Performance Keywords

parallelize               → Convert sequential fetches to Promise.all
waterfall                 → Find and fix request waterfalls
bundle                    → Analyze bundle size, suggest splits
lazy                      → Add dynamic imports for heavy components
suspense                  → Add Suspense boundaries

Autopilot Mode

Run with --dangerously-skip-permissions for uninterrupted 100-turn cycles:

alias c='claude --dangerously-skip-permissions'
c "build the complete settings page with profile, notifications, and billing tabs"

Claude will:

  1. Plan the approach
  2. Create components
  3. Connect data layer
  4. Add i18n keys
  5. Test at 3 breakpoints
  6. Fix issues automatically

Remote Control

Every Claude Code session on the host Mac registers with Remote Control, so the Claude app's Code tab reaches it directly:

Mac:    Start feature work → it gets long → walk away
Phone:  Push arrives when it finishes or needs a decision (only while you're away from the Mac)
Phone:  Code tab → open the session → answer, steer, /model, /effort, attach a photo
Phone:  Or start something new: Code tab → "Mac · kun" (the always-on kun-rc server)
  • Done with a session? Close it. The kun-rc server is the way back in, so nothing has to stay open "in case". Each idle session holds ~10 MCP processes on a 16 GB Mac.
  • Mac going off? Start a cloud session from the Code tab instead; it runs on GitHub repos and keeps going with the lid closed.
  • Check the lane: bash ~/.claude/scripts/devices.sh --status (disk, swap, sessions, push, presence, server).

Agent Teams (Experimental)

Parallelize feature development:

"Build settings: profile tab, notifications tab, billing tab"
→ Lead Agent dispatches:
  ├── Agent A (worktree) → Profile form
  ├── Agent B (worktree) → Notifications
  └── Agent C (worktree) → Billing + Stripe
→ Lead merges → single PR

Shell Aliases

# Quick start
alias c='claude --dangerously-skip-permissions'
 
# Resume last session
alias cr='claude -r'
 
# Direct prompt
alias cq='claude -p'