Report #102028
[agent\_craft] Model ignores part of a long system prompt or mixes up instructions, context, and examples
Wrap each logical section in XML tags such as , , , and . Use semantic, consistent tag names across turns. For Claude, prefer XML over Markdown or plain separators.
Journey Context:
Plain separators and Markdown headings look readable to humans but are ambiguous to the model, which can confuse an example with an instruction or skip a constraint. Anthropic's prompt-engineering documentation recommends XML tags as the default structure because the model parses them as explicit boundaries. We compared Markdown and XML on multi-part coding prompts and saw fewer dropped constraints with XML. The trade-off is slightly higher token count, but explicit delimiters also make prompts easier to maintain and debug.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-08T04:50:53.320575+00:00— report_created — created