Harper Reed's LLM Codegen Workflow

Overview

TL;DR: Brainstorm spec → plan a plan → execute using LLM codegen in discrete loops

Two Development Scenarios

Greenfield Development

Brand new projects starting from scratch

Legacy/Modern Code

Iterating on existing codebases

Greenfield Process

Step 1: Idea Honing

• Use conversational LLM (ChatGPT 4o/o3) for iterative brainstorming
• Ask for one question at a time to develop thorough specifications
• End with comprehensive developer-ready specification
• Save as spec.md

Key Prompt: > "Ask me one question at a time so we can develop a thorough, step-by-step spec for this idea..."

Step 2: Planning

• Pass spec to reasoning model (o1, o3, r1)
• Create detailed step-by-step blueprint
• Break down into small, iterative chunks
• Generate series of prompts for code-generation LLM
• Save as prompt_plan.md and todo.md

Two Planning Approaches:

• TDD Prompt: Test-driven development focused
• Non-TDD Prompt: Standard implementation approach

Time Investment: ~15 minutes total

Step 3: Execution

Claude Method:

• Pair programming with claude.ai
• Manual copy/paste workflow
• Use repomix for debugging when stuck
• Author handles initial boilerplate and tooling setup

Aider Method:

• Automated "cookie clicker" experience
• Hands-off testing and debugging
• Direct prompt execution

Workflow Steps:

1. Set up repo boilerplate
2. Execute prompts iteratively
3. Test and verify functionality
4. Debug if needed
5. Repeat until completion

Non-Greenfield Process

Context Generation

• Use repomix tool to package codebase
• Generate output.txt with code context
• Utilize mise task runner for automation

Available LLM Tasks:

• Clean bundles
• Copy buffer bundle
• Generate code reviews
• Generate GitHub issues
• Generate missing tests
• Generate README

Execution Methods:

Same Claude and Aider approaches as Greenfield, but task-focused rather than project-wide planning

Utility Prompts

Code Review

"You are a senior developer. Your job is to do a thorough code review..."

GitHub Issue Generation

"Review this code, and write out the top issues you see..."

Missing Tests

"Review this code, and write out a list of missing test cases..."

Key Challenges & Solutions

"Getting Over Your Skis"

• Problem: Easy to lose track and get ahead of yourself
• Solution: Aggressive tracking and planning documentation
• Recommendation: Take breaks when overwhelmed

Solo Development Limitation

• Current workflows are single-player experiences
• Team collaboration with LLMs remains problematic
• Merges and context sharing are complicated

Downtime Management

While waiting for LLM token processing:

• Start brainstorming next project
• Listen to music
• Play games
• Social interaction

Results & Benefits

Productivity Gains:

• Built numerous projects: scripts, apps, CLI tools
• Works across multiple programming languages
• Empty hack to-do list due to rapid execution
• Exploration of new programming languages and tools

Time Efficiency:

Planning phase takes only ~15 minutes, enabling rapid project initiation and completion