# AIP016: claude-specific-tags-on-gpt
## Description
Detects Anthropic-specific XML tags (`<thinking>`, `<scratchpad>`, `<reflection>`, etc.) in prompts when targeting GPT models. These tags are Anthropic conventions and are ignored by GPT models, making them ineffective.
## Severity
Warning
## Applies When
This rule only fires when `target_models` is configured to include models with names starting with `gpt-` (case-insensitive) or exactly `"openai"`. If no target models are specified, this rule produces no diagnostics.
## Example
```
model: gpt-4
prompt: |
Please reason inside <thinking>...</thinking> before answering.
```
Violation: The `<thinking>` tag is not supported by GPT models and will be treated as literal text.
## Remediation
Replace Anthropic-specific tags with model-agnostic techniques:
- Use standard JSON/YAML schemas for structured reasoning.
- Employ explicit step-by-step instructions instead of hidden tags.
- Test prompts on actual target models before deployment.