Here’s how I use LLMs to help me write code

Takeaways

• LLM-assisted coding is difficult and unintuitive — treat models as an over-confident, fast pair programmer, not AGI; augment your skill, don’t abdicate it.
• Context is king: most craft is managing what enters the conversation; reset threads when they go stale; prefer tools that expose context clearly.
• Account for training cut-offs when picking libraries; favor boring, stable, well-represented stacks or feed recent examples explicitly.
• Start projects by asking for options and prototypes before committing to an implementation path; iterate simple → sophisticated within one context.
• Don’t anthropomorphize failures — note tasks models can’t do; a stronger model is one that suddenly handles a previously impossible task.
• Provide full working examples in prompts; use conversation history deliberately; test and review everything the model produces.

Notes

Simon Willison’s foundational practitioner guide on LLM coding — sets expectations and patterns that complement Boris Tane’s research-plan workflow and Matt Pocock’s plan loop.

Open questions

• Which of Willison’s context-management habits should be encoded as vault agent rules vs left to session practice?