gemini-prompts

Instruction file imported from Dayanara847/arborschool-content (.cursor/rules/gemini-prompts.mdc). Copyright stays with the author.

Gemini Prompting Rules

When you create or modify prompts intended for Gemini (code, docs, or Cursor chat instructions), follow docs/gemini-3-pro-prompt-engineering-best-practices.md.

Current model: We use gemini-3.1-pro-preview.

• Be precise and direct

  • Avoid over-engineered, verbose prompts.
  • State the task in 1–3 clear sentences.

• Structure prompts

  • Use sections like Role, Task, Rules, Output Format, Context, Final Instruction as recommended in the doc.

• Control output format

  • Be explicit about output format (JSON/XML/text) and verbosity.
  • Note: thinking_level may not be directly supported by the SDK adapter.

• Anchor to context

  • For large inputs, put context first, instructions last, and use an anchor phrase such as:
    "Based on the information above, …".

• Avoid overfitting prompts to specific examples

  • When a prompt fails on one input, fix the underlying principle, not the symptom.
  • Prefer simplifying existing rules over adding new exception rules.
  • Maintain a separate evaluation set to verify changes generalize.
  • Red flags: rules referencing specific IDs/values, accumulating exceptions, frequent prompt changes after each failure.