Ai/ProgramLoop
Skills
- /WritePRD
Flow
Idea + /GrillMe
- /WritePRD
- Kanban tasks , want vertical tracer bullet slices
- Prompt: The first task/slice is to horizontal
- Use prd skill to write local markdown plans/issues .md tasks
- Let ai loose in containers to exploit kanban parallelism using TDD red/green
- Want deep modules not lots of shallow modules
Sandcastle - typescript parallel runner using GitHub workspaces and containers
Grill me skill
- The Grill Me skill text:
- "Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies one by one. For each question, provide your recommended answer. Ask the questions one at a time."
- Key purpose:
- It prevents misalignment by forcing the AI to probe into the details of the requested task (12:34).
It ensures the user and the AI have a shared design concept before any coding begins (16:23).
- It prevents the AI from simply guessing or rushing to create a plan before it truly understands the requirements (16:36).
Summarise the video
Summarize the bullet points for developer using GitHub copilot and mcp
Based on the walkthrough provided in the video, here are the key strategies and workflows for developers leveraging AI agents like GitHub Copilot (and similar coding agents) to ship production features effectively:
- Core Development Lifecycle:
- Stress-test requirements:
- Instead of just writing code from vague briefs, start by turning ambiguous requirements into structured Product Requirements Documents (PRDs) using interviewing techniques like the "Grill Me" skill to ensure alignment between the human and the AI (12:17 - 17:35).
- Stress-test requirements:
Vertical slicing: Break down large tasks into "tracer bullet" vertical slices—thin pieces of functionality that cross all layers (schema, service, front end) rather than horizontal slices that only focus on one layer at a time (43:32 - 45:15). TDD (Test-Driven Development): Utilize the AI to write tests first, implement the code, and then run feedback loops to ensure the system is integrated and working correctly (56:15 - 57:30). Agent Management & Optimization:
The Smart Zone: Maintain a lean system prompt to keep your context window optimized; avoid overloading the AI with unnecessary data that pushes it into the "dumb zone" (7:50 - 8:15). Human-in-the-loop vs. AFK: Start with a human-in-the-loop approach to refine prompts and workflows, then gradually move toward autonomous (AFK) runs once you have a reliable backlog and clear definitions of done (52:26 - 55:35). Sandboxing: Run agents in isolated environments (e.g., Docker sandboxes) to safely test code commits without compromising the local development environment (55:20 - 55:35). Architecting for AI Effectiveness:
Deep Modules: Design codebases with deep, well-defined modules. Complicated or fragmented architectures make it difficult for both human developers and AI agents to work effectively (113:42 - 120:00). Avoid Documentation Rot: Do not over-index on maintaining legacy PRDs. Once a task is complete, it is often better to archive or close the associated documentation to prevent future agents from being misled by outdated plans (123:30 - 125:05).
What is the "dumb zone"? Why does Matt prefer human-led PRD creation? How should code reviews be handled with agents? Give me the GrillMe skill text The Grill Me skill text, as described by Matt Pocock in the video (16:17), is designed to be a short, direct prompt for the AI to ensure alignment during the planning phase.
Based on the walkthrough provided in the video, here are the key strategies and workflows for developers leveraging AI agents like GitHub Copilot (and similar coding agents) to ship production features effectively:
Core Development Lifecycle:
Stress-test requirements: Instead of just writing code from vague briefs, start by turning ambiguous requirements into structured Product Requirements Documents (PRDs) using interviewing techniques like the "Grill Me" skill to ensure alignment between the human and the AI (12:17 - 17:35). Vertical slicing: Break down large tasks into "tracer bullet" vertical slices—thin pieces of functionality that cross all layers (schema, service, front end) rather than horizontal slices that only focus on one layer at a time (43:32 - 45:15). TDD (Test-Driven Development): Utilize the AI to write tests first, implement the code, and then run feedback loops to ensure the system is integrated and working correctly (56:15 - 57:30). Agent Management & Optimization:
The Smart Zone: Maintain a lean system prompt to keep your context window optimized; avoid overloading the AI with unnecessary data that pushes it into the "dumb zone" (7:50 - 8:15). Human-in-the-loop vs. AFK: Start with a human-in-the-loop approach to refine prompts and workflows, then gradually move toward autonomous (AFK) runs once you have a reliable backlog and clear definitions of done (52:26 - 55:35). Sandboxing: Run agents in isolated environments (e.g., Docker sandboxes) to safely test code commits without compromising the local development environment (55:20 - 55:35). Architecting for AI Effectiveness:
Deep Modules: Design codebases with deep, well-defined modules. Complicated or fragmented architectures make it difficult for both human developers and AI agents to work effectively (113:42 - 120:00). Avoid Documentation Rot: Do not over-index on maintaining legacy PRDs. Once a task is complete, it is often better to archive or close the associated documentation to prevent future agents from being misled by outdated plans (123:30 - 125:05).