How I Make Claude Code Cook: A 4-Phase AI Development Workflow
Source: Video transcript analysis
Author's Background: Building with AI exclusively for several years
Executive Summary
This workflow represents a mature, production-tested approach to AI-assisted development. The author emphasizes that the Define phase represents 70% of meaningful work and is the primary determinant of project success. The workflow scales from individual coding tasks to complex research and file transformation operations.
Phase 1: DEFINE (The Critical 70%)
Core Principle
"If I am successful in a project, it is almost certain I'm successful because of this phase."
The Capture Process
1. Voice Recording / Walking Ideation
- Free-associate ideas verbally while walking
- Don't aim for final conclusions during ideation
- Allow mind to meander and explore possibilities
- Raw ideas are valid—refinement comes later
2. Visual Documentation
- Create annotated screenshots of current state
- Mark up images showing desired changes
- Visual notes complement verbal ideas
3. PRD Generation (Using ChatGPT or similar)
- Combine voice transcript + visual notes + component definitions
- Critical constraint: Focus on WHAT, not HOW
- Zero technical implementation details
- Document should tell the story of user experience
PRD Content Requirements
| Element | Description |
|---|---|
| User capabilities | What should users be able to see/do? |
| Restrictions | What should users NOT be able to see/do? |
| Interactions | How do parts relate to each other? |
| Priorities | What's most important? |
| Emotional outcome | What should the user feel? |
Example PRD Language
"The end feeling should be: 'I grab the window, I move it, everything listens to me, nothing fights back.'"
Phase 2: REFINE (Iterative Improvement)
Context Management (Critical)
The Problem: Context window accumulation degrades model performance.
The Solution: Clear context aggressively.
/clear # Create fresh chat context
/context # Visualize current context window
Rule of Thumb: Clear context whenever you finish a discrete task, or when the next task isn't directly related to the previous conversation.
Plan Mode Workflow
- Use
Shift+Tabto cycle through modes - Enter Plan Mode before implementation
- Plan Mode = conversational, no file writes
- Iterate on plan until satisfied ("measure 50 times")
- Reference PRD and visual assets during planning
The CLAUDE.md File
This persistent memory file loads into every conversation. Configure for:
- Code style preferences (clean code, specific patterns)
- Testing requirements (unit tests for critical logic)
- Universal behaviors you want in every session
Screenshot Integration
Why it matters: Models cannot see your screen. They infer UI state from code.
Workflow:
- Use screenshot utility with quick keystroke (CleanShot, native tools)
- Annotate screenshots with arrows, highlights, notes
- Paste directly into terminal with
Ctrl+V - Reference specific visual elements: "Do you see the two pink arrows?"
Testing Discipline
Reality Check: Models know how to write tests but won't remember to do so unprompted.
Required Reminders:
- Explicitly request test case creation
- Remind to run test cases
- Add testing requirements to CLAUDE.md
- Consider: unit tests, visual tests, E2E tests, Playwright integration
Consequence of Neglect:
"If you don't have good tests, three months down the road, you will cry."
Phase 3: MULTIPLY (Parallel Agent Execution)
The Unlock
Use Claude AI web interface (claude.ai) or equivalent platforms (Codex, Gemini) to run multiple agents in parallel.
Capabilities
| Feature | Benefit |
|---|---|
| Parallel execution | Multiple changes simultaneously |
| Asynchronous operation | Not tied to terminal/machine |
| Cloud-based | Works from phone, survives shutdown |
| Exploratory | Kick off 4-5 variants, pick best approach |
Workflow Pattern
- Navigate to claude.ai → Projects
- Select active project (same codebase as local)
- Submit change requests as natural language
- Multiple requests run simultaneously
- Review completed PRs when ready
Use Cases
- Exploring multiple implementation directions
- Parallelizing independent features
- Batch processing similar changes
- Disconnected development (mobile, travel)
Phase 4: EXPAND (Beyond Code)
PR Management via Claude Code
# Check open PRs
"Hey, what are the open PRs?"
# Review and merge
"I have two PRs open against this repo. I've verified both.
Approve them, merge them back, and update main."
Result: Complete GitHub workflow without leaving terminal.
Custom Slash Commands
Power Feature: Create project-specific commands that persist with the repository.
/git-push # Custom push workflow
/test # Project-specific test runner
/dev # Development environment setup
/install # Installation sequence
Benefits:
- Check into version control
- Available when returning to project weeks later
- Encode project-specific knowledge
File Transformation (Non-Code Operations)
The Revelation: Claude Code works on any local files, not just source code.
Example: JSON Data Enhancement
- Input: Board game history file with short descriptions, broken images
- Process: 45-minute autonomous deep research
- Output: Validated images, extended descriptions, verified external links
Other Demonstrated Uses:
- Audio file conversion
- Image analysis and renaming
- Document research and compilation
- Meeting notes processing
- Video script writing
The Harness Model
Claude Code operates as an iterative problem solver:
- Creates internal to-do list
- Works through items systematically
- Can run unattended for extended periods
- Different from "thinking models" that make internal passes
Key Tactical Patterns
Context Commands
| Command | Effect |
|---|---|
/clear | Fresh chat context |
/context | Visualize context window |
Shift+Tab | Cycle through modes |
Custom /commands | Project-specific automation |
Mode Selection
| Mode | Use Case |
|---|---|
| Plan Mode | Conversational planning, no writes |
| Implementation Mode | Active file creation/modification |
Success Factors Summary
- Define Phase Investment — 70% of success determined here
- Context Hygiene — Clear frequently, minimize window size
- Visual Communication — Screenshots with annotations
- Testing Discipline — Explicit reminders, persistent config
- Parallel Execution — Multiply agents for exploration
- Slash Commands — Encode project knowledge
- File Transformation — Extend beyond code to any file type
Workflow Diagram
┌─────────────────────────────────────────────────────────────────┐
│ DEFINE (70%) │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Voice │ + │ Visual │ + │Component │ → │ PRD │ │
│ │ Notes │ │ Notes │ │ Docs │ │(WHAT not │ │
│ │(Walking) │ │(Annotate)│ │ │ │ HOW) │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ REFINE │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │Plan Mode │ → │Iterate on│ → │Screenshot│ → │ Write │ │
│ │(Measure │ │ Plan │ │ Debug │ │ Tests │ │
│ │50 times) │ │ │ │ │ │ │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
│ ↑ │
│ /clear frequently │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ MULTIPLY │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ Claude AI Web Interface │ │
│ │ ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐ │ │
│ │ │Agent 1 │ │Agent 2 │ │Agent 3 │ │Agent N │ │ │
│ │ │(Async) │ │(Async) │ │(Async) │ │(Async) │ │ │
│ │ └────────┘ └────────┘ └────────┘ └────────┘ │ │
│ └─────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ EXPAND │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────────────┐ │
│ │PR Management │ │Slash Commands│ │ File Transformation │ │
│ │via Terminal │ │(Persistent) │ │ (Any local files) │ │
│ └──────────────┘ └──────────────┘ └──────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Author's Final Assessment
"If you took a few of these away, I honestly would not be as successful as I feel like I am."
The workflow represents years of refinement and demonstrates that AI development tools have matured beyond simple code generation into comprehensive development lifecycle management.