Grok Bot 工程师用技能文件把信任建起来
一位参与 Grok Bot 的 SpaceX AI 工程师从首个 bug 发现代理没读代码就下结论，于是把缺失上下文写成 Markdown 技能，五个月后单月 1000 个 PR。
A SpaceX AI engineer who helped ship Grok Bot discovered something important from her very first bug. “Oh, it has to be this.” The agent sounded completely certain. But its tool calls revealed the truth: it had never even opened the code it was blaming. Most people would call the model unreliable and return to checking every line manually. She took a different approach. Instead of blaming the model, she turned the missing context into a reusable skill: a simple Markdown file explaining how the agent should handle the task. Five months later: → 1,000 PRs in a single month → Agents reviewing and auto-merging their own work The process was simple: Prompt → Tool calls → Missing context → Skill → Trust The problem wasn't necessarily the model. It was the missing context that kept it from doing the job right. Bookmark this before blaming your next AI agent for getting your code wrong.