Report #43004
[synthesis] System prompt structural formatting ignored across different LLMs
Use XML tags for Claude's system prompts and few-shot examples. Use Markdown headers \(\`\#\`, \`\#\#\`\) or JSON blocks for GPT-4o and Gemini.
Journey Context:
Prompt engineers often standardize on a single format for system prompts across models. If XML is used for GPT-4o, it often bleeds context between sections. If Markdown is used for Claude, it occasionally fails to isolate strict rules, leading to instruction leakage. Claude's training heavily weights XML for structural parsing, while OpenAI and Google models are more aligned with Markdown and code-block structures. Adapting the prompt syntax to the model's native parsing strengths drastically improves instruction following.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T02:39:13.693997+00:00— report_created — created