Individual sentences carry limited info, but 85% accuracy per sentence means we can be highly confident about longer texts. This far exceeded my expectations. Old-school alchemy still rules—way better than those dumb online tools that just ask an LLM, “Is this text AI-generated?”
Deny list anxiety. Claude Code’s permission system has evaluation order issues where blanket Bash allows override of the ask list. This unfortunate bug has led me to do away with the ask list and instead use a deny list. I deny destructive commands (rm, git reset --hard, DROP) but the agent keeps finding creative ways to override them: unlink, python -c "import os; os.remove()", find ... -delete. I recently added an instruction in CLAUDE.md to not do that, so far so good but I don’t fully trust it.
。业内人士推荐搜狗输入法下载作为进阶阅读
В Финляндии захотели пойти на опасный шаг против России02:50
Like Ponomari, she says it can be "disappointing" when audiences never see a scene you've worked on, but agrees it's "just part of the game".