Some observations from this past month.
Context
- When building context docs (RAG, Skills, Rules, etc.), minimise the number of “don’t do this” examples. During long-running conversations, the LLM sometimes mixes up the DO’s and the DON’TS.
- RAG, Skills, etc., are often ignored in favour of training data. As an example, if you’re making a web framework similar to React and providing data via RAG, Skills and Tools, there is a chance the agent will pattern match with React from its training and go with it.
- This might make you consider using fine-tuning. This should be your last resort when using 3rd party models.
- Should you use MCP or give the model console access? Do both.
Coding
- Codex is great at code reviewing Claude Code’s output.
- Claude Code’s code when fixing bugs can err towards the simple and least intrusive fix. You have to ask it to be thorough if you want the most stable fix.
- The Agent gave you a couple of implementation options and its recommendation, and you have no idea what the best approach actually is? Ask it what the “unwanted consequences” would be so you can go into a deeper conversation with it.
- LLMs are great at TypeScript but less good at languages like Rust. Niche languages like COBOL fare even worse.
GenAI
- GenAI image generation remains bad. Use image…