[FULL WORKSHOP] AI Coding For Real Engineers - Matt Pocock, AI Hero (@mattpocockuk )

Key Concepts

• Smart Zone vs. Dumb Zone: The theory that LLMs perform optimally with smaller context windows (the "Smart Zone," roughly <100k tokens). As context grows, attention relationships scale quadratically, leading to degraded performance (the "Dumb Zone").
• Memento Principle: Treating LLM sessions as ephemeral, stateless interactions that reset to a base state, rather than relying on long, bloated conversation histories.
• Grill Me Skill: A methodology for aligning with an AI by having it act as an interviewer, asking relentless questions to reach a "shared design concept" before any code is written.
• Tracer Bullets (Vertical Slices): A development strategy where functionality is implemented in thin, end-to-end slices (crossing database, API, and UI layers) to ensure immediate feedback, rather than horizontal, layer-by-layer development.
• Deep vs. Shallow Modules: Based on John Ousterhout’s A Philosophy of Software Design, deep modules expose simple interfaces but contain significant functionality, making them easier to test and maintain compared to shallow, fragmented modules.
• AFK (Away From Keyboard) Tasks: Tasks that can be delegated to an AI agent to run autonomously, provided the human has established a clear destination and feedback loop.

1. The Development Framework

The speaker argues that software engineering fundamentals are more critical than ever when working with AI. The workflow is structured into three distinct phases:

• Alignment (Human-in-the-loop): Using the "Grill Me" skill to interview the AI and define the "design concept." This prevents the "specs-to-code" trap where developers ignore the codebase.
• Destination & Journey Planning: Creating a Product Requirements Document (PRD) to define the end state, followed by a Kanban board to break the work into independent, parallelizable issues.
• Implementation (AFK): Running autonomous agents (using tools like Claude Code or custom scripts like Sand Castle) to execute tasks, utilizing TDD (Test-Driven Development) to ensure quality.

2. Methodologies and Processes

• The "Grill Me" Process: Instead of writing a static spec, the developer provides a brief and instructs the AI to "interview me relentlessly." This forces the developer to clarify assumptions and reach a shared understanding.
• Compacting vs. Clearing: While some developers "compact" (summarize) long histories to save tokens, the speaker prefers clearing the context entirely to maintain a consistent, predictable base state.
• TDD (Red-Green-Refactor): The speaker mandates TDD for AI agents. By forcing the agent to write a failing test first, the developer ensures the AI isn't "cheating" or writing code that lacks proper integration.
• Parallelization: By structuring tasks as a Directed Acyclic Graph (DAG) on a Kanban board, multiple agents can work on independent issues simultaneously, significantly speeding up development.

3. Key Arguments

• Codebase Quality is Paramount: "Bad codebases make bad agents." If the architecture is shallow or poorly structured, the AI will struggle to navigate dependencies.
• Avoid "Specs-to-Code": The speaker strongly criticizes the "vibe coding" movement where developers ignore the code and only edit specs. He argues that the code is the "battleground" and must be actively shaped.
• Human Taste is Irreplaceable: Automation should be applied to implementation, but the "taste" and high-level architectural decisions must remain human-led. Over-automating the entire lifecycle leads to "slop."

4. Notable Quotes

• "I needed to reach a shared understanding. I didn't need an asset, I didn't need a plan; I needed to be on the same wavelength as the AI."
• "If you're getting bad outputs from your AI, you often need to increase the quality of your feedback loops."
• "Democracy is the worst way to run a country apart from all the other ways—that's how I feel about Claude Code."

5. Synthesis and Conclusion

The main takeaway is that AI should not be treated as a magic box that replaces the developer, but as a tool that requires rigorous engineering discipline. By keeping tasks small (within the "Smart Zone"), enforcing vertical slices for feedback, and maintaining deep module architectures, developers can effectively delegate the "night shift" (implementation) to AI while retaining control over the "day shift" (design and alignment). The speaker emphasizes that the best way to improve AI output is to improve the quality of the feedback loops (tests and types) within the codebase itself.