On Magic Words
Or, how to make your tacit expertise useful for your AI work
This past week I had the good fortune to speak with a group of smart historians at the University of Virginia about the uses of AI for history. Few of them were coders, but all of them wanted to learn how to better use AI.
I did spend a little time talking about coding and its many uses, like reproducibility and deterministic computing, but they were really interested in how to make the best use of agents, like Claude. During the conversation, as I was teaching, I realized something implicit that I need to make more explicit: magic words.
Magic words are those special incantations dropped into a conversation with an AI that make explicit something you implicitly know it should do.
For instance: “check your work.” You ask the AI to do a task and it accomplishes it with all the naïve enthusiasm and unwarranted confidence of an inexperienced intern, and with all the frustrations therein.
The difference is that the AI does the task in seconds, unlike a human intern, so you mistakenly assume it knows what it is doing. Often, it does not. If you were doing the work, you would double-check your work. Neither the intern nor the AI will do so automatically. “Check your work” makes the implicit explicit. Tell an AI to “check your work,” and suddenly it is making sure all the numbers add up properly and all the i’s are dotted. And unlike a real person, you can make that super-fast AI check its work instantly. And unlike the real world, you can just have another dozen AI’s check the work of the first AI for nearly no cost in time or money.
In a thousand different ways, I find that using AI is very similar to teaching. Experts forget how much invisible scaffolding they carry around. As a teacher, I try— successfully or unsuccessfully—to reattain beginner’s mind and remember what I had to learn to do a task.
In my classroom, I try to make all my tacit knowledge of working with archives and making historical arguments into clear instructions. I find this extremely challenging because “the right way” is now deeply engrained in how I see the world.
This problem is well known in educational circles. It is why a graduate student often has an easier time explaining historical argumentation (“now you signpost with an analytic topic sentence in this way…”) than I do (“write the correct thing in the correct way that obviously makes more sense”). I can remember what it was like to have the graduate student’s intermediate control of the discipline, but can not longer experience it easily.
Magic words are all about reclaiming that tacit expert knowledge. Here are some from my own spellbook.
“You are a manager. Spawn an appropriate model subagent to do the tasks. Run in parallel.”
One of the superpowers of agents like Claude is that they can make more of themselves, like the brooms from The Sorcerer’s Apprentice in Fantasia or Mr. Meeseeks from Rick and Morty.
Ask Claude to be a manager and to spawn subagents to do its tasks. Ask it to use the correct model—dumb, average, or smart—so you do not waste tokens. Run agents in parallel (at the same time) and your agents will rapidly speed up your work.
“Spawn an adversarial AI agent and check the work.”
Whenever I do something non-trivial, I ask Claude to spawn an “adversarial AI” to critically examine what has happened. Sometimes I spawn multiple agents for different parts of a program (“make sure the data contracts are followed”, “check for edge cases”) or an essay I have written (“read for structural logic,” “check the footnote citations,” “check the footnote format”). The adversarial check is key to vibe coding. AI makes lots and lots of confident mistakes. You cannot and will not catch them all. Use the AI to check and correct the AI.
During the workshop, I realized that I had lots of tacit knowledge about how to interact with AI. Some examples are below:
“Iterate until it works correctly. Spawn adversarial AI to check after each step.”
“Run experiments using different models to find the best solution.”
“Read this source to confirm, deny, or complicate my argument.”
“Keep track of all bugs and solutions in solutions.md.”
“Keep a log of all decisions and key findings in logbook.md.”
“How would you prompt another AI to solve this problem?”
“Give me your plan and let’s discuss it before doing anything.”
“What should I be asking you to do this well? What assumptions am I making?”
The breakthrough was not any particular set of magic words, but the realization that they existed at all—and that I already knew them. You do too. This tacit knowledge of what works and doesn’t work is how you move from beginner to expert. We use magic words all the time with our students, and now you should be using them to make the best use of AI.
Please share your magic words in the comments!
The next time you teach, think about the magic words you use with your students. I suspect these are the same spells that will empower you in your use of AI in your own work.
Great post, Louis, it certainly can seem like 'magic' when one has developed the (often painful) experience over innumerable interactions with tools like Claude Code that can be boiled down to simple, effective prompts. Agent Skills.md are one of the main mediums I use for making tacit knowledge explicit, for codifying procedural knowledge, which is the crucial prerequisite to automation. I highly encourage others to read more about the Skills framework, which is a fascinating glimpse of the future of human-language programming: https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview As Anthropic's engineers noted, we can start to think of the LLM as the processor, agents as the operating system, and skills as applications or software. Who could be better suited to this new, human-language-based paradigm of computing than humanities scholars, the masters of language?