Prompting remains crucial despite stronger LLMs, but requires a different skill set for agents like Claude Code. Good prompts are explicit, include testing instructions, and clarify ambiguities, while bad prompts are vague and make unfounded assumptions. Prompting is also an iterative process of working with the model to plan implementations.

I use Fluid Voice, a free local transcription tool, to input prompts. Speaking is two to three times faster than typing and reduces hand strain. More importantly, transcription makes it effortless to include all necessary context, which is essential for the agent to solve tasks effectively.

Always tell the model how to test implementations. For a chat feature, instruct it to open Chrome, send a query, and verify the AI response and token streaming. This saves significant manual testing time. Additionally, have a discussion with the agent before implementation. Ask it to research, check logs, and surface ambiguities or significant assumptions — such as choosing AWS CLI versus CDK for infrastructure — to align your mental model with the agent's plan.

HTML reports can document these discussions and decisions for future reference.