Use workflows to spin up separate "worker" and "verifier" sub-agents. A model is more lenient checking its own work ("self-preferential bias"). Using a distinct agent with a clear rubric for verification leads to higher quality, more reliable outputs for non-deterministic tasks.
Effective AI planning isn't a one-shot command. It's an iterative exploration to understand system limitations, edge cases, and what you actually want. Use the agent to generate explainers (e.g., on Whisper's failure modes) to eliminate blind spots before committing to a complex workflow.
As AI models become more capable, overly detailed system prompts with many examples and hard constraints can be counterproductive. They limit the model's creativity. The Claude Code team cut their system prompt by 80% because the smarter model needs more freedom to find optimal solutions.
The team treats Slack as the "multiplayer" environment for AI. Collaborative tasks, PR management, and background processes ("multi-clotting") are handled via Claude Tag in shared channels. The local terminal (Claude Code) is reserved for a single, focused, iterative task, preventing context switching overload.
Instead of using absolute negatives like "never do X," explain the underlying reason you want to avoid X. This gives the model flexibility. For example, rather than a hard character limit, explaining the goal is a single tweet but allowing a thread if necessary gives the AI more freedom to create a better output.
Learning to work effectively with AI agents isn't about memorizing code syntax, which the agent handles. True technical skill is developing a deep understanding of the underlying systems' constraints and tradeoffs (e.g., different video libraries) to better guide the agent toward optimal solutions.
Traditional specs are static. Instruct the AI agent to maintain "implementation notes" as it works. This captures unexpected issues and discoveries made during development. The spec then evolves from a starting point into a living document that informs future iterations and potential respecs.
Before committing to a full implementation in a framework like React, use an AI agent to generate simple HTML artifacts that explore different design variations. This is a fast and cheap way to prototype visual concepts, especially for non-designers who "know it when they see it."