Full Walkthrough: How Does Matt Pocock Structure an AI Coding Workflow?

TL;DR

Matt Pocock’s AI coding workflow keeps tasks small enough for an LLM’s “smart zone,” then supports implementation with structured planning, vertical task slices, tests, feedback loops, and human QA. He identifies roughly 100k tokens as a practical marker where performance starts declining, regardless of whether the context window is 200k or 1 million tokens. Read on for the specific planning and engineering practices behind the workflow.

Transcript

Yeah, we good. >> Okay, folks, we're at capacity. Let's kick off. I don't want you waiting here for 25 more minutes before we some arbitrary deadline. So, welcome. My name is Matt. Uh I'm a teacher and I suppose now I teach AI. Um we have a link up here if you've not already been to this which is has the exercises for the um stuff we're going to do... Read More

Key Insights

• AI can transform ambiguous requirements into structured, actionable plans.
• Vertical slicing of tasks ensures feedback at each layer of development.
• Test-driven development (TDD) is essential for AI to produce reliable code.
• Deep module design in codebases enhances testability and AI efficiency.
• Human oversight is crucial in planning and quality assurance stages.
• AI agents can perform coding tasks autonomously with proper setup.
• Feedback loops are critical for AI to produce high-quality outputs.
• AI coding workflows require a balance between automation and human input.

Questions & Answers

Q: How does Matt Pocock structure an efficient AI coding workflow?

Matt Pocock recommends turning ambiguous requirements into a structured plan and dividing large jobs into small tasks that remain within the LLM’s “smart zone.” Vertical slices, test-driven development, feedback loops, deep modules, and human review then help the agent implement each task reliably.

Q: What are the LLM “smart zone” and “dumb zone”?

The “smart zone,” an idea attributed to Dex Hy of Human Layer, is the early part of a conversation where attention relationships are least strained and the LLM does its best work. As more tokens enter the same context, the model moves toward the “dumb zone” and begins making poorer decisions.

Q: When does an LLM begin leaving its smart zone?

Matt gives roughly 100k tokens as his current marker for when an LLM starts getting less capable. He says this occurs whether the advertised context window is 200k or 1 million tokens.

Q: How should large coding tasks be divided for AI agents?

Break a large task into small, independently manageable sections so each can be completed within a fresh, effective context. Matt describes multi-phase plans as a common starting point and suggests treating repeated phases as a loop of smaller tasks.

Q: What is the role of vertical slicing in AI coding?

Vertical slicing breaks work into thin tasks that cross the necessary application layers. Each slice supplies feedback while preserving integration and functionality across those layers.

Q: Why is test-driven development important for AI agents?

Test-driven development gives an AI agent an explicit feedback cycle: write a failing test, implement the required behavior, and refactor. This structure helps verify that generated code satisfies the stated requirements.

Q: How does codebase design affect AI agent effectiveness?

Deep modules encapsulate functionality behind simple interfaces, reducing the complexity an AI agent must handle. This design also improves testability, making it easier for the agent to navigate and modify the codebase.

Q: Why is human oversight still needed in an AI coding workflow?

Humans remain responsible for clarifying requirements, shaping plans and prompts, and checking whether the finished work meets project goals. QA and code review provide oversight even when an agent performs implementation autonomously.

Summary & Key Takeaways

• AI can convert vague requirements into structured plans, enabling autonomous agents to implement features efficiently. This involves vertical task slicing, test-driven development, and designing codebases that support AI operations. Human oversight is essential in planning and QA stages to ensure quality.

• AI agents can work autonomously with a well-structured workflow, but human input remains crucial for refining prompts and ensuring the final product meets standards. The integration of AI into coding requires a balance between automation and human oversight.

• By designing codebases with deep modules and employing test-driven development, AI agents can operate more effectively. This approach maximizes agent efficiency and ensures the production of high-quality code, while human oversight maintains control over the development process.