General

Be explicit & literal: Spell everything out exactly as you want it
Induce planning/CoT: +4% gain. Explicitly tell the model to plan/think step-by-step if needed.
Put instructions before and after long context: If you can only choose one, choose before
- Kind of makes sense - the attention mechanism needs to know what to look for; asking it after having provided the context goes against how I understand attention
Use XML or Markdown, not JSON
Structure prompts: Use sections (Role, Instructions, Examples, etc.) using XML or Markdown (see above)
Check for contradicting rules: Later rules win; but contradicting instructions hurt performance
Not necessary to use threats or ALL CAPS if the rules in this post are followed

Agents

Agent Reminders (persist, use tools, plan): ~20-24% gain. Tell agents to keep going, use their tools (don't guess), and optionally plan out loud.
Use API tools field, not prompts: +2% gain. Define tools properly in the API, don't just describe them in the text.
Specify context reliance: Explicitly instruct whether to only use provided context or allow mixing with internal knowledge/tool calling.

Performance may degrade if many items need retrieval or complex reasoning spans the entire context.
Resistance to long repetitive outputs
Parallel tool calls can sometimes fail