Wink Pings

AI Is Smarter Than You—So Why Are You Still Teaching It What to Do?

Lauren teaches AI how to do things; Baoyu lets AI learn on its own. Set the goal and acceptance criteria, and leave the rest for the model to run. What's the difference between these two automation approaches?

Lauren Tan wrote an introductory guide to automation. Its core idea is to first clarify the goal, then teach the bot how to do it. She mentions a key insight: automation is not just technical work; it is a way of thinking. Break the goal into steps, then let the machine run them for you. The significance of Grok Bot is that people who don't know how to code can use natural language to get a bot to do the work.

Baoyu reposted this article and said he does it differently. The difference is this: Lauren is very capable, and the models are not yet smart enough, so she spends her effort teaching the Agent, handing her own best practices to the Grok bot so it can think like she does. Baoyu believes AI is smarter than he is, and the models he uses are not limited to Grok, so he emphasizes goals and acceptance criteria and does not interfere with the intermediate steps.

His method is a loop: goal → acceptance → improvement.

There isn't much to say about the goal. Acceptance criteria tell AI how to verify the results itself, ideally with tools: screenshots, web search, and a code execution environment. For example, when he automated news publishing, an important acceptance criterion was that fact-checking must be done before writing, to ensure the content is true and sourced. Some tasks can also produce a test set or automated test code, so that every run is automatically verified.

The third step is to get it running first, then adjust. Baoyu's analogy is target shooting: take a shot at the target first; after the shot, you know how far you are from the bullseye and which way you are off, so you have a clear idea of how to adjust next. As for how the Agent carries out the work, you don't need to look very closely. Adjusting the results is where professional judgment is most tested. You yourself must know what good looks like; otherwise you can only accept whatever AI generates. It's not enough to see that something is bad—you also need to be able to accurately tell AI where to change it.

The fourth step is to let it evolve on its own. Today's Agent Harness has memory, but it does not necessarily proactively record problems encountered in automation tasks. He improves the Agent on two levels: first, distilling best practices into a Skill; second, having the Agent proactively record lessons learned during execution, then go back and improve the Skill itself. When he works on reverse-engineering tasks, after each round of execution he has the Agent write an entry in lessons.md—write down any lesson learned, skip it if there is none. Then he lets the model evaluate for itself whether to optimize the Skill's instructions and scripts.

The side effect is that the Skill becomes increasingly bloated, so it needs regular refactoring and cleanup, and this is left to the Agent itself. After each refactor, rerun the previous test sets and automated tests to avoid deleting something by mistake.

There were several angles in the comments. One person joked that lessons.md sounds like the Agent's diary, dedicated to recording the days when it did poorly. Another raised a more practical problem: the more best practices you hand over, the more you need to guard against rule drift—rules that have already been retired may still be treated as valid in the next round. This is the hardest kind of drift in multi-round handoff experiments. Someone else summed it up more bluntly: lock down the goal and acceptance criteria, let go of the intermediate steps, and once the loop is spinning, it's more cost-effective than teaching step by step.

Another commenter added context: Agent Skills was officially handed over to the Linux Foundation last month, and more than 70 platforms, including Claude Code, Codex, and ChatGPT, now read the same SKILL.md. lessons.md plus regular refactoring means you write it once and can reuse it across platforms. Others raised a counter-condition: when model resources are limited, without more human constraints the results are unusable; this is essentially a trade-off, not a case of one method being more advanced.

At the end of the article, Lauren included a Dr Eggbot v2 bot template designed to help people generate automation bots with a single responsibility; it first interviews them about the goal and then helps them build it. She said many people don't use Grok Bot to its full potential, because knowing how to automate is itself a skill that has to be learned.

One commenter nailed the difference between the two approaches: she is still writing the lecture notes, while he has already gone to collect the exam papers. Neither is absolutely right. The smarter the model, the less effort it takes to set the goal and acceptance criteria and let AI run on its own. If AI is already smarter than you, teaching it what to do step by step is indeed a bit of a loss.

发布时间: 2026-11-20 15:36