Key Points

CLI-based AI

  • CLI agents coordinate actions across multiple files.
  • Run /init inside Gemini to create a GEMINI.md Living Spec that reduces context drift.
  • A portable AGENTS.md lets the same spec travel across different AI tools.
  • Research orchestration shifts focus from writing syntax to validating intent.

Best practices for prompting

  • Be specific and provide context.
  • Plan before you act: request a numbered plan and approve it before any files are written.
  • Always validate AI outputs.
  • Introspection improves code quality.

Data cleaning with AI

  • Use AI to automate data harmonization and standardization.
  • Audit data before cleaning to identify inconsistencies.
  • Read and test all generated code before running it.

Validation strategies: the approval gate

  • The approval gate separates experimental prototypes from validated research.
  • Rewrite time is the primary metric for measuring AI workflow value.
  • Requirement constraints prevent the AI from drifting away from research specs.
  • Multi-model verification uses a second AI to act as a skeptical peer reviewer.

Limitations and cautions

  • Avoid AI for security-critical tasks.
  • You are responsible for the final output.
  • Open models offer better reproducibility; proprietary models offer more power.

Resources and next steps

  • Gemini, Claude, and Copilot serve different needs.
  • Community support is vital.
  • Plugins and MCP allow AI to connect to external data.