2026-10-11 17:09 UTC

prompt-engineering

band: coolmomentum: stable score: 0.175
temperature history

Episodes (2)

Anthropic's official Opus 5.5 prompting guide documents harness-breaking mechanics — thinking can no longer be disabled, progress updates arrive as often-empty thinking blocks, and some turns end with plain text instead of tool calls that unattended agent loops mistake for task completion — forcing coding-agent harnesses to adopt the guide's mitigation patterns to keep Opus 5.5 agents running.
corroboratedconvergesscott: high
OpenAI claims GPT-6 Astra needs shorter, selectively loaded skills and task-specific instructions with explicit completion boundaries, making legacy instruction-heavy Codex configurations a source of wasted context, unnecessary testing, and premature stopping.
watchingconvergesscott: medium