The 5 Techniques Separating Top Agentic Engineers Right Now

How to Maximize Your AI Coding Assistant's Potential

Introduction to AI Coding Assistants

• The speaker emphasizes that many users are not fully utilizing the potential of their AI coding assistants, such as Claude, Code Hero, or Cursor.

• Assumes a basic understanding of coding agents and aims to provide practical techniques used by experienced engineers.

Importance of PRD (Product Requirement Document)

• Defines PRD as a markdown document outlining the entire scope of work for a project, crucial for both greenfield and brownfield development.

• For greenfield projects, the PRD includes everything needed for proof of concept or MVP; it serves as a guiding document for feature development.

• Highlights the importance of breaking down projects into granular features to avoid overwhelming the coding agent.

Creating and Utilizing PRDs

• In brownfield development, the focus shifts to documenting existing codebases and planning future developments while maintaining a clear project direction.

• A GitHub repository is mentioned where users can find demo projects and commands related to creating effective PRDs.

Workflow Commands for AI Development

• Introduces core slash commands used daily in AI coding workflows, including one specifically designed for creating PRDs.

• Describes how to engage with an AI assistant by discussing project goals before generating a comprehensive PRD using the /create PRD command.

Structuring Feature Development Around PRDs

• Once the PRD is established, it becomes essential for guiding subsequent feature development with the help of an AI coding assistant.

Priming Your Coding Agent

• Discusses using a "prime" command at the start of conversations with an AI assistant to load necessary context from the project files like the PRD.

Modular Rules Architecture

• Introduces modular rules architecture as a way to keep global rules concise and relevant across different tasks without overwhelming the LLM (Large Language Model).

Best Practices in Rule Management

• Advises keeping global rule files short and focused on universal constraints while splitting task-specific rules into separate documents.

Example: Habit Tracker Application Rules

• Provides insight into managing rules effectively through examples from a habit tracker application, emphasizing lightweight rule files.

Understanding Project Structure for Coding Agents

Importance of Project Context

• The speaker emphasizes the need for coding agents to understand project structure, including commands for front-end and back-end operations, server types, code conventions, and logging standards.

• A reference section is introduced where task-specific context is loaded only when necessary, allowing the coding agent to access relevant information based on the feature being developed.

Managing Reference Information

• The reference folder in the codebase contains extensive documentation (nearly a thousand lines), which provides detailed instructions specific to API development.

• Keeping global rules concise while maintaining access to comprehensive context is crucial; this helps protect the coding agent's context window from becoming overloaded.

Commandification of Processes

• The speaker advocates for "commandifying" repetitive prompts sent to the coding agent, suggesting that any prompt used more than twice should be converted into a reusable command or workflow.

• By documenting core commands in markdown format within a repository, developers can streamline their workflows and save time during development tasks like git commits and code reviews.

Comprehensive Workflow Documentation

• The speaker shares that all essential commands are documented in a repository for others to customize according to their needs. This includes various aspects of feature development cycles.

• Techniques covered include executing planning phases, validating processes, and evolving systems over time with AI assistance.

Context Management Techniques

• A critical technique discussed is resetting context between planning and execution phases by outputting a document containing all necessary information before starting code writing.

• This approach ensures minimal context load during execution allows the coding agent ample room for reasoning and self-validation without unnecessary distractions.

Execution Planning Process

• During execution planning, no additional priming is done; instead, a structured plan document serves as the sole input for the coding agent.

• The process involves using commands like /cle to clear previous contexts before executing new plans based solely on previously defined documents.

This structured approach aims at enhancing efficiency in working with AI-driven coding agents while ensuring clarity and focus throughout development processes.

System Evolution and Coding Agents

The Importance of System Improvement

• Treating bugs as opportunities for enhancement can significantly strengthen coding agents. Instead of merely fixing issues, developers should analyze the system to prevent future occurrences.

• Identifying patterns in recurring issues allows for targeted improvements in global rules or workflows, enhancing the overall reliability of the coding agent.

Strategies for Addressing Bugs

• When a coding agent encounters an error, it often indicates a misunderstanding of rules or validation processes that need refinement.

• Examples include adding new rules for import styles or updating templates to ensure testing is included in execution plans.

Continuous Learning and Adaptation

• After completing feature development, it's crucial to review any discrepancies between expected outcomes and actual results. This reflection helps identify areas for improvement.

• Developers should engage their coding agents in self-reflection by comparing execution against established plans and rules to address bugs effectively.

Mindset Shift: Fixing Systems Over Bugs

• Emphasizing a mindset where developers focus on improving systems rather than just fixing individual bugs leads to more powerful and reliable coding agents over time.

• Techniques shared are commonly used by top engineers, highlighting the importance of systematic thinking in software development.