developing-ai-prompts

<essential_principles>

<principle name="be_explicit"> State exactly what you want. Models follow instructions precisely—if you want elaborate output, ask for it explicitly. "Create an analytics dashboard with as many relevant features as possible" outperforms "Create an analytics dashboard." </principle>

<principle name="add_context"> Explain WHY behind instructions. "Never use ellipses because the text-to-speech engine won't know how to pronounce them" beats "NEVER use ellipses." Models generalize from explanations. </principle>

<principle name="usexmlstructure"> Wrap distinct sections in XML tags (<instructions>, <examples>, <context>). This creates unambiguous boundaries, enables selective attention, and improves parsing accuracy across all models. </principle>

<principle name="showdonttell"> Few-shot examples outperform lengthy explanations. Show 2-3 input/output pairs demonstrating the exact format and style you want. Models learn patterns better from demonstrations than descriptions. </principle>

<principle name="positiveovernegative"> Tell the model what TO do, not what NOT to do. "Write in flowing prose paragraphs" beats "Don't use bullet points." Positive patterns guide behavior more effectively than prohibitions. </principle>

<principle name="matchspecificityto_stakes"> High-stakes operations (security, payments, data migrations) need exact instructions with no freedom. Creative tasks (writing, analysis, code review) benefit from high freedom with principles over procedures. </principle>

<principle name="review_holistically"> Prompts edited by multiple people or composed from multiple sources degrade over time. Each engineer adds emphasis for their concern, implicitly diluting others. Periodically review the entire prompt as a system: remove duplicates, tighten language, rebalance intent. A holistic view beats "prompting harder." </principle>

</essential_principles>

<intake> What would you like to do?

1. Write a new prompt or system instruction
2. Review/improve an existing prompt
3. Adapt a prompt for a different model
4. Get guidance on a specific technique

Wait for response before proceeding. </intake>

<routing>

Intent-based routing:

• "system prompt for Claude" → workflows/write-prompt.md + references/claude-specific.md
• "why isn't my prompt working" → workflows/review-prompt.md
• "convert GPT prompt to Gemini" → workflows/adapt-for-model.md
• "how do I use few-shot" → references/universal-techniques.md

</routing>

<quick_reference>

Universal Structure:

<role>Who the model is</role>
<context>Background information</context>
<instructions>What to do</instructions>
<examples>
<example>
<input>Sample input</input>
<output>Expected output</output>
</example>
</examples>
<constraints>Boundaries and limitations</constraints>
<output_format>How to structure response</output_format>

Model Selection:

• Claude: Best for nuanced reasoning, long context, agentic tasks. Responds well to XML, context explanations, thinking blocks.
• GPT: Strong instruction adherence, scope discipline. Benefits from explicit length constraints, schema definitions.
• Gemini: Excellent multimodal, strong few-shot learning. Use input/output prefixes, positive examples only.

</quick_reference>

<reference_index> All in references/:

Techniques: universal-techniques.md (XML structure, few-shot, chain-of-thought, output formatting) Claude: claude-specific.md (Claude 4.x best practices, thinking, tool use, agentic patterns) GPT: gpt-specific.md (GPT-5 patterns, verbosity control, scope discipline) Gemini: gemini-specific.md (Gemini 3 patterns, prefixes, temperature, multimodal) Avoid: anti-patterns.md (common mistakes across all models) </reference_index>

<workflows_index> All in workflows/:

</workflows_index>

<success_criteria> A well-crafted prompt:

• States the task explicitly with clear success criteria
• Provides context explaining why instructions matter
• Uses XML structure for distinct sections
• Includes 2-3 few-shot examples when format matters
• Matches specificity to task stakes
• Works reliably across multiple test cases
• Avoids known anti-patterns for the target model

</success_criteria>