Prompting Foundation Patterns
Abstract
Principled prompt designs (role/task setup, exemplars, chain-of-thought style hints) that consistently improve reliability across tasks.
Motivation
- Reduce variance from underspecified instructions
- Enable decomposition via step-by-step prompting
Architectures
- System/role priming + task instruction + constraints + examples
- CoT-style scaffolding (don’t reveal chain-of-thought in outputs for safety)
- ReAct-like text-only: reasoning tokens with action descriptions sans tool execution
Design Choices
- Few-shot vs. zero-shot with meta-instructions
- Output constraints (bullets, JSON, tables) vs. free-form
- Guardrails: banned topics, safety reminders
Pros/Cons
- Pros: More reliable outputs, better controllability
- Cons: Longer prompts increase cost; risk of leaking internal rubrics
Evaluation Metrics
- Task success rate, rubric adherence
- Format correctness; post-edit rates
Vendor/Tooling
- Provider system messages; structured outputs for schemas
- Prompt caching and templates in SDKs
Design Checklist
- State role, goal, success criteria
- Provide canonical examples; specify format
- Include safety and refusal guidance
- Constrain decoding and validate outputs
References
- Title: Best practices for prompting (OpenAI) URL: https://platform.openai.com/docs/guides/prompt-engineering Publisher/Vendor: OpenAI Accessed: 2025-08-14 Version_or_release: provider_reported
- Title: Anthropic Prompting Guides URL: https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering Publisher/Vendor: Anthropic Accessed: 2025-08-14 Version_or_release: provider_reported