Software Engineering :: Pair Programming :: AI Assistant :: RepoPrompt :: About :: The Secret AI Prompt Tool Silicon Valley Engineers Are Using
⪼ Made with 💜 by Polyglot.
This podcast interview between Nathan Lands (host of The Next Wave) and Eric Provenger (creator of Repopromp) explores the shifting landscape of AI-assisted coding. The episode aims to educate developers—especially those beyond “vibe coding”—on how to get serious, high-quality results from LLMs by providing explicit, structured context. Eric introduces Repopromp as a powerful tool for selectively feeding relevant code into LLMs using code maps, token-aware file selection, and structured prompts (especially in XML). They also discuss the evolving role of software engineers in an AI-first world and the trade-offs between open-sourcing, VC funding, and solo development.
- Even Andrei Karpathy now stresses the need for structured context instead of relying on "vibe coding"
- LLMs perform significantly better with curated, relevant input
- Repopromp lets developers handpick and prioritize files for LLMs
- Lets you select and organize relevant parts of a codebase for LLM input
- Uses a "code map" (index-like abstraction) to compress and summarize large codebases
- Helps you structure context using XML so models like Claude or GPT can parse and act effectively
- Effective use of context windows (8k, 32k, up to 1M tokens) is essential for model performance
- Selective context curation gets better results than flooding models with entire codebases
- LLM output improves with smart XML formatting versus raw JSON or natural language
- Assigning LLMs explicit roles (e.g., “Architect” vs. “Engineer”) improves output relevance
- Iterating on prompts after reviewing output is key to refinement
- Use structured “plans” for modifications instead of relying on vague instructions
- Academics and legal researchers are adopting Repopromp for non-code text workflows
- Future plans include support for MCP (Model Context Protocol) to enhance integration with external tools
- Eric built Repopromp solo during paternity leave and continues development on weekends
- Open sourcing comes with risks—projects like Roo have already overtaken Klein via forks
- VC funding may not align with the current vision; building for the passionate userbase remains the focus
- The role of engineers is shifting toward context curation, architecture, and guiding models
- New coders may lack the foundational struggle that seasoned developers faced
- Over-reliance on AI tools can lead to skill atrophy—intentional struggle may be necessary for true learning
- Built natively (Swift) for performance reasons—Electron created too many cross-platform headaches
- Designed for large-scale parallel file processing not optimized in traditional IDEs
- Plans to enhance agent support, parallel experimentation, and real-time documentation retrieval
- Belief that AI will handle more work, but context-building remains a critical human role
- Sees potential in open protocols like MCP to expand tool ecosystem compatibility