How to Build an Effective AI Coding Workflow

TL;DR

Improve AI coding results by optimizing the harness around the model: prompts, skills, documentation, tests, task scope, interfaces, and codebase design. AI can handle much of tactical programming, but developers still need strategic programming skills to design hard parts, delegate clearly, evaluate the work, and create an environment where agents can make effective changes.

Transcript

Everyone's obsessed with the uh model and I think they should be more interested in the harness, what you can do to get the most out of the harness, giving it the right prompts, giving it the right skills to work with and improving the environment in which the model runs. As I sort of said with Fable, like the model is useful, but I think the harne... Read More

Key Insights

• The harness is as important as the model because developers can directly improve prompts, skills, documentation, tests, and the environment in which AI operates. Focusing only on a new model overlooks the practical systems that determine whether an agent can make accurate and efficient changes.
• Tactical programming is the day-to-day work of writing code, handling syntax, investigating bugs, and creating commits. AI can perform much of this work more cheaply, which shifts the developer’s role toward directing, structuring, and evaluating the work rather than manually completing every implementation step.
• Strategic programming is the long-term work of designing a codebase, choosing approaches that improve velocity, and coordinating how separate parts fit together. Developers who strengthen these abilities can make better use of AI agents because they can assign clearer work and recognize whether the resulting implementation is sound.
• Good AI delegation requires the hard parts to be designed upfront and each task to have a narrow, explicit scope. The same principles used when delegating to junior or mid-level programmers still apply when an AI agent becomes the implementer.
• Codebase structure affects both AI performance and token spending because an easier-to-change codebase requires less effort to understand and modify. Clear module interfaces, useful tests, and sufficient documentation give agents richer context and point them toward the correct locations for changes.
• Human expertise sets the ceiling for useful AI output because skilled practitioners can provide richer context, stronger direction, and better oversight. The transcript argues that AI gives senior developers a particularly large boost, while people with limited domain knowledge receive a smaller benefit.
• The teach skill encodes educational principles such as the zone of proximal development and distinctions among knowledge, skills, and wisdom. It uses those principles to create a course dynamically, and Pocock reports using it to learn to solve a Rubik’s cube from memory.
• Mission alignment is the first stage of the demonstrated teaching workflow because learning is intended to help someone accomplish something concrete. The skill asks what the learner is building and what better software means in their situation before deciding which gaps and subjects deserve priority.

Questions & Answers

Q: How can developers improve an AI coding workflow?

Developers can improve an AI coding workflow by working on the harness surrounding the model. That includes writing better prompts, providing appropriate skills, maintaining clear documentation, creating useful tests, defining module interfaces, and making the codebase easier to change. They should also design difficult parts upfront and divide implementation into well-scoped tasks that an AI agent can complete and verify.

Q: What is the difference between tactical and strategic programming?

Tactical programming covers immediate implementation work such as writing code, working through syntax, fixing bugs as they appear, and creating commits. Strategic programming focuses on longer-term outcomes, including codebase design, development velocity, task organization, and relationships between modules. The transcript argues that AI now handles much of the tactical work, making strategic ability increasingly important for developers directing agents.

Q: Why is the AI harness as important as the model?

The harness determines how the model receives context, instructions, capabilities, and feedback while performing work. Developers have substantial control over prompts, skills, documentation, tests, and the surrounding codebase, even when they cannot control the model itself. Improving these elements can help an agent locate the correct code, understand the requested change, work within clear boundaries, and produce a more effective result.

Q: What skills are needed to delegate coding tasks to AI agents?

Effective delegation requires developers to design the difficult parts before implementation, scope tasks carefully, define interfaces between modules, prepare useful tests, and maintain enough documentation to guide an agent. These practices are not presented as new AI-specific principles. They are the same strategic skills used when delegating work to junior or mid-level programmers, with AI taking the role of the tactical implementer.

Q: Why does codebase quality affect AI token spending?

A codebase that is easier to understand and modify gives an AI agent clearer paths to the relevant files, interfaces, and behaviors. The transcript directly connects lower token spending with having a codebase that is easier to change. Clear organization, tests, documentation, and module boundaries reduce the amount of investigation needed before an agent can make an effective modification.

Q: How does domain expertise affect results from AI?

Domain expertise acts as a multiplier because a knowledgeable person can supply richer context, specify better solutions, and oversee the resulting work more effectively. According to Pocock, a person’s skills are the ceiling on what AI can do for them. A skilled teacher can use AI to teach better, just as a skilled developer can direct AI to build and modify software more effectively.

Q: What does Matt Pocock’s teach skill do?

The teach skill applies principles drawn from Pocock’s teaching experience, including the zone of proximal development and distinctions among knowledge, skills, and wisdom. It can investigate trusted resources, assemble a curriculum, and create a course dynamically around a learner’s goal. Pocock says he used it to teach himself to solve a Rubik’s cube from memory and to explore becoming a senior developer.

Q: What should a beginner or vibe coder learn after basic syntax?

The demonstrated teaching workflow suggests that the highest-leverage gap is often not additional syntax. It identifies surrounding engineering abilities such as Git, reading errors, debugging, understanding how software ships, and testing. Before choosing a curriculum, however, it asks about the learner’s concrete mission, current project, and intended meaning of shipping better software so the lessons support a practical outcome.

Summary & Key Takeaways

• The central argument is that developers focus too heavily on choosing the newest AI model. They have more control over the harness, including prompts, skills, documentation, tests, and the codebase itself. Improving that environment helps AI find the correct places, understand intended changes, and complete work more effectively.

• AI has largely absorbed tactical programming, including writing code, handling syntax, investigating bugs, and creating commits. Human advantage increasingly comes from strategic programming: designing difficult parts, defining module interfaces, planning for long-term velocity, and breaking work into well-scoped tasks that an effectively unlimited group of AI programmers can execute.

• Domain expertise determines how effectively someone can direct and evaluate AI. Matt Pocock demonstrates this idea with a teaching skill that applies principles from his teaching experience, investigates trusted resources, creates a curriculum, and stores information in a workspace. It begins by clarifying the learner’s mission instead of immediately delivering generic information.