Claude Code setup: prompts, skills and evaluation checks

This Claude Code tutorial presents AI Labs' interpretation of workflow advice from creator Boris Cherny. It focuses on reviewing a coding assistant's setup after model changes: remove instructions the model can infer from the project, then retain unusual requirements and recurring mistakes.

The speaker describes using the usage and context commands to inspect overhead, safe mode to try tasks without custom configuration, and the doctor command to propose cleanup. These are the video's reported behaviors. Its claims about model improvements and resistance to prompt injection do not establish universal reliability or safety.

A central recommendation is to define completion checks before work starts. Examples include rejecting an incorrect password and checking whether a page fits a phone screen. The speaker recounts a Swift rewrite of the Claude desktop app that used screenshot comparisons, while stressing that people still need to decide what correct means.

For instructions, the proposed order is a clearer prompt, a short Claude.md for persistent project requirements, skills for occasional detail, and MCP when information is otherwise inaccessible. Prompts should state the desired result and constraints. The closing section cautions against large dynamic agent workflows and describes routines as recurring cloud jobs that continue after the laptop closes.