1137 Checklist 7165eb4b

Prompt Optimization Checklist

Universal Criteria (All Platforms)

Structure (20 points)

• Clear section headers or XML tags
• Logical flow: context → instructions → examples
• No mixed formatting (pick one style)
• References linked appropriately for platform

Specificity (20 points)

• Instructions use numbered steps
• Success criteria explicit
• Concrete examples, not abstract
• Edge cases addressed

Examples (20 points)

• 3-5 diverse examples present
• Examples relevant to actual tasks
• Include edge cases
• Show good AND bad examples

Conciseness (15 points)

• No redundant explanations
• Trusts model knows basics
• Essential info only
• Tables over paragraphs where appropriate

Model Awareness (15 points)

• Model-specific clauses added
• Appropriate guidance level for target model
• No over-explaining for smart models

System/User Separation (10 points)

• Persistent behavior in system prompt
• Task-specific in user prompt
• No confusion between roles

Platform-Specific Checks

Claude API

• Uses XML tags for structure
• Tools in API tools field (not prompt)
• Opus: Has anti-over-engineering clause
• Haiku: Has sufficient examples (3-5)
• Avoids "think" word if extended thinking disabled

OpenAI API

• Uses message roles correctly
• Model version pinned for production
• Tools in tools parameter
• Verbosity controlled explicitly

Gemini API

• Single format (XML OR Markdown)
• Critical instructions at END for long context
• Temperature not modified (keep 1.0)
• Knowledge cutoff mentioned if relevant
• Grounding clause if using context

CLAUDE.md

• Content inline (not progressive disclosure)
• Under 200 instructions
• Includes: context, principles, workflows
• Excludes: API docs, rarely-used details

ChatGPT Custom Instructions

• Under 1500 characters per field
• Clear "About me" section
• Clear "How to respond" section
• Works with Memory feature

n8n / Workflow Tools

• Dynamic variables for dates/context
• Tools named clearly
• Memory node configured
• Explicit boundaries defined

Anti-Pattern Check

Scoring

Interpretation