Programming with AI: You're Probably Doing It Wrong
16 days ago
- #Model Context Protocols
- #LLM best practices
- #AI-assisted development
- 2025 is the year of AI, with GPT-5 prompting developers to reassess their use of LLMs in daily work.
- Common mistakes in AI-assisted development include manually providing context, shipping unverified code, skipping testing, and treating AI as smart auto-complete.
- Proper AI usage requires providing the right context and specific instructions (prompts) to guide the LLM effectively.
- Agents are specialized prompts that define how an LLM should behave in specific contexts, such as front-end design or software architecture.
- Model Context Protocols (MCPs) act like USB connectors for AI, providing specific context and tasks, such as accessing databases or GitHub projects.
- Examples of MCPs include postgres-mcp for database access and github-mcp-server for GitHub project integration.
- Caveats include the effort required to write proper agents, potential risks with MCP permissions, and the need to manage agent definitions and MCPs.
- AI does not replace software engineers but shifts their role to overseeing AI operations and stepping in when issues arise.