Conference/ORCwAI/Pt2/1320 Lightning Talks
1320 lightning talks
Modern Day Mashups: How AI Agents Are Reviving the Programmable Web – Person/Angie Jones (5 minutes) (5 minutes)
- very cool talk about Goose MCP/Client and how we're in a new mashup generation
Action items
Conference/ORCwAI/Pt2/1320 Lightning Talks/p2 Tipping AI Code Generation on its Side – Person/Craig McLuckie (5 minutes)
Summary
2025-05-08 – O'Reilly AI Codecon – Craig McLuckie – Tipping AI Code Generation on its Side
Talk Overview
- Speaker: Craig McLuckie
- Length: 5 minutes (Lightning Talk)
- Time: 13:24 ET
- Theme: Moving from closed, vertically integrated AI tooling to open, horizontally composable systems
Slide: Vertically Integrated AI Stacks
- Visual shows 3 identical vertical stacks, each labeled App at the top.
- Each stack contains:
- Workflow Logic
- Context Management
- LLM
- Represents current architecture where every AI tool builds its own full vertical system.
- This leads to duplication of infrastructure across apps.
AI Summary
- Craig McLuckie critiques the prevailing model of AI code generation, where each product builds an end-to-end vertical stack—including its own LLM, context manager, and workflow engine.
- These vertically integrated stacks are closed systems that inhibit interoperability, innovation, and reuse.
- He argues that the next evolution must tip this structure horizontally, enabling shared context management and LLM resources across different applications.
- This shift mirrors earlier evolutions in tech (e.g. microservices, containerization) that replaced monolithic systems with more modular, composable architectures.
- Emphasizes the importance of community-led efforts and open standards—like the Model Context Protocol (MCP)—to facilitate this transition.
Conference/ORCwAI/Pt2/1320 Lightning Talks/p3 Prompt Engineering as a Core Dev Skill Patty OCallaghan – Person/Patty OCallaghan - Techniques for Getting High-Quality Code from LLMs
slides
2025-05-08 – O'Reilly AI Codecon – Patty O’Callaghan – Techniques for Getting High-Quality Code from LLMs
Slide Summary
- Title Slide
- "Techniques for Getting High-Quality Code from LLMs"
- Slide: Prompt Engineering is a Core Dev Skill
- LLMs are now tools in every developer’s toolbox.
- Prompt engineering is as fundamental as learning a framework.
- Slide: Architecture-First Prompting
- Begin prompts with high-level architecture and design goals.
- Helps guide LLMs to generate code that fits into existing systems.
- Example prompt: “I’m building a Flask app with this structure…”
- Slide: Contextual Scaffolding
- Provide surrounding code context to situate the LLM.
- Use file headers, imports, docstrings, and nearby function calls.
- Helps the LLM understand naming conventions and system patterns.
- Slide: Task-Specific Prompting
- Customize prompts for:
- New feature development
- Debugging
- Refactoring
- Example: For debugging, provide error trace and expected behavior.
- Slide: Progressive Enhancement
- Start with a minimal viable prompt.
- Add detail iteratively to refine the output.
- Encourages experimentation and layered improvement.
- Slide: Collaborate, Don’t Delegate
- Treat the LLM like a junior developer:
- Guide it.
- Review output.
- Improve over time.
- Emphasizes the importance of human review and iteration.
AI Summary
- Patty O’Callaghan presents a set of practical strategies for getting consistently strong output from LLMs in software engineering tasks.
- Key idea: don’t expect magic—collaborate with the model by designing thoughtful prompts.
- Strong prompts lead to better alignment with existing systems and more maintainable code.
- Encourages iterative development of prompts, tailored to both the task and the architecture.