walkthrough.md

The ideal workflow for building with Mrs. Kitty

This isn’t a typical AI coding tool. It thinks before it acts. Here’s how to work with it.

The Big Picture

Think architect, not bricklayer.

You’re building a house. You design the blueprint — plumbing, electricity, walls — measured from the top down. You don’t need to know which brick goes on top of which brick. Claude handles the bricks. The system frees you from prompt engineering minutiae so you can focus on design and architecture. That’s where the full creative potential opens up — because the engineering is handled.
Key insight: The less you micromanage the code, the better the output. Describe what you want at the highest level of abstraction you can. The system breaks it down from there.
Step 1

Describe Your Vision

  • Start by describing what you want to build
  • Think it out. Give as much context as possible
  • The more context upfront, the fewer iterations later
$ claude
> I want to build a task management app with...
  - User authentication with Google OAuth
  - Kanban board with drag-and-drop
  - Real-time updates via WebSockets
  - Mobile responsive design
Tip: Don’t worry about technical specifics yet. Describe the experience you want. “Users should be able to drag tasks between columns and see changes instantly” is better than “use React DnD with Socket.io”.
Step 2

Answer the Questions

  • The system asks clarifying questions automatically
  • Multiple-choice, one question at a time
  • It’s gathering context to build the right thing
┌─ Question ───────────────────────────────────────┐
│ Which database should we use?
│
│  ● PostgreSQL (Recommended)
│  ○ MongoDB
│  ○ SQLite
│  ○ Other
└────────────────────────────────────────────┘
Tip: Answer honestly. If you don’t know, say “I don’t know” — it will research for you.
Step 3

Review the Plan

  • It enters plan mode automatically
  • Researches: web searches, documentation, codebase exploration
  • Writes a step-by-step implementation plan
  • Read the plan. If too technical, ask Claude to summarize in plain language
  • This is your last checkpoint before code starts
✓ Entered plan mode
  Exploring codebase... 3 agents launched
  WebSearch: "best OAuth implementation 2026"
  WebFetch: next-auth.js documentation
  Reading: src/auth/config.ts

  Plan written to plan file (14 steps)
  Ready for your approval.
Important: You need to understand what’s about to be built. If you don’t understand it, don’t approve it. Ask questions until you do.
Step 4

Watch It Build

  • Approve and implementation begins
  • TDD loop: writes tests first, implements, verifies
  • Sub-agents spawn for complex tasks — normal behavior
  ▸ Launching kraken agent... implementing auth
  ▸ Launching arbiter agent... writing tests
  ✓ 12/14 steps complete
  ✓ All tests passing (34/34)
  ▸ Browser verification: snapshot verified (200 tokens)
Tip: Let it cook. For complex tasks, it spawns background agents that work in parallel. This is intentional — each agent gets clean context.
Step 5

Steer the Ship

  • As it builds, interject with design changes, new features, characteristics
  • This is where the creative potential opens up
  • Don’t be 100% reliant on AI — share your expertise and direction
  • The ship works, but you make it the best possible build
Tip: Add functions, change the design, adjust the architecture — in real time. You’re the creative director. The system is your team.

What You’ll See

System behaviors that might surprise new users.

“Error” Messages That Aren’t Errors

Red “error” text appears from hooks even when everything worked correctly. This is a known behavior in Claude Code’s plugin system — hooks report via the error channel even for success messages. Anthropic hasn’t fixed this yet. It’s cosmetic, not functional.

Rule of thumb: If work continues normally after the “error”, it’s fine.

⚠ PreToolUse:Edit hook error: Edit applied   ← This is SUCCESS
  File: walkthrough.html                         ← The edit worked
  Backup: (version_manager)                     ← Backup was made

  [Claude continues working normally...]     ← No problem!

Gate Blocks — Messages That Pause Progress

  • “Read these files before proceeding” — Project Analyzer learning your codebase
  • “Confidence: 72%. Research needed.” — Thinking Partner ensuring quality (needs 97%)
  • “Enter plan mode” — Requires your approval before code changes
  • “Attempt 2 of 3” — Tiered warning (3 chances before hard block)

These aren’t errors — they’re the system being careful.

Maintenance Mode — Bypass Gates When Needed

  • Temporarily disables all gates for quick changes
  • When: Infrastructure fixes, config tweaks, emergencies
  • How: Claude will ask you — say “enable maintenance mode”
  • Auto-expires after 1 hour. Always re-enables when done.

Memory System — “I found relevant memories...”

  • Remembers across conversations
  • Recalls what worked, what failed, decisions you made
  • Ask “what did we do last time with [topic]?” to search manually

Auto Research — Web Searches Happen Automatically

  • Searches docs, best practices, existing solutions before implementing
  • You’ll see WebSearch/WebFetch calls — that’s the research gate

Reflexion — Intelligent Failure Recovery

  • 1 failure — researches why before retrying
  • 2+ failures — searches from at least 4 different angles
  • Not brute force — learns from mistakes

Browser Automation — For UI Work & Beyond

  • Uses your real Chrome — 140+ commands for clicking, filling, navigating, and verifying. Completely undetectable
  • Reads page structure in ~200 tokens instead of screenshots — 98.5% less context than Playwright

Sub-Agents — “Launching scout agent...”

  • Complex tasks spawn specialists (researcher, implementer, tester)
  • They work in separate contexts — main conversation stays clean
  • Normal and intentional — context quality protection

Power Features

Shortcuts and capabilities for faster workflows.

Voice Dictation

Alt+X

Press to start recording, press again to stop. Your speech is transcribed and sent as a message. Hands-free coding.

Screenshots

Win+Shift+S then Alt+V

Capture any part of your screen, then paste it directly into the conversation. Show Claude what you see.

Slash Commands

/commit /fix /build /explore /review

Pre-built workflows for common operations. Type the command and let the system handle the rest.

Multi-Terminal

Ctrl+Shift+T

Open new terminal tabs. Sessions auto-coordinate via database — they know about each other and avoid file conflicts.

Fresh Start

cn

Type cn to start a new conversation. Clean context, fresh start. Memory carries over automatically.

Session Management

/clear /compact /cost /context

Clear history, compress long conversations, check token costs, or see context usage. Type /help for the full list.

Ready to build?

Download Mrs. Kitty and start building with the workflow above.

Download Free