“Unusual was obliged to invent a lot of the magic he did, working from normal rules and half-remembered tales from previous books.”
— Susanna Clarke, Jonathan Unusual & Mr Norrell
Fairy tales, myths, and fantasy fiction are stuffed with magic spells. You say “abracadabra” and one thing profound occurs.1 Say “open sesame” and the door swings open.
It seems that that is additionally a helpful metaphor for what occurs with giant language fashions.
I first acquired this concept from David Griffiths’s O’Reilly course on utilizing AI to spice up your productiveness. He gave a easy instance. You possibly can inform ChatGPT “Manage my process listing utilizing the Eisenhower four-sided field.” And it simply is aware of what to do, even if you happen to your self know nothing about Common Dwight D. Eisenhower’s method to determination making. David then suggests his college students as an alternative attempt “Manage my process listing utilizing Getting Issues Completed,” or simply “Use GTD.” Every of these phrases is shorthand for methods of thought, practices, and conventions that the mannequin has discovered from human tradition.
These are magic phrases. They’re magic not as a result of they do one thing unworldly and sudden however as a result of they’ve the ability to summon patterns which were encoded within the mannequin. The phrases act as keys, unlocking context and even whole workflows.
All of us use magic phrases in our prompts. We are saying one thing like “Replace my resume” or “Draft a Substack submit” with out pondering how a lot detailed prompting we’d should do to create that output if the LLM didn’t already know the magic phrase.
Each area has a specialised language whose phrases are recognized solely to its initiates. We may be fanciful and fake they’re magic spells, however the actuality is that every of them is mostly a form of fuzzy perform name to an LLM, bringing in a physique of context and unlocking a set of behaviors and capabilities. Once we ask an LLM to write down a program in Javascript fairly than Python, we’re utilizing considered one of these fuzzy perform calls. Once we ask for output as an .md file, we’re doing the identical. Not like a perform name in a standard programming language, it doesn’t all the time return the identical outcome, which is why builders have a chance to boost the magic.
From Prompts to Purposes
The following gentle bulb went off for me in a dialog with Claire Vo, the creator of an AI software known as ChatPRD. Claire spent years as a product supervisor, and as quickly as ChatGPT grew to become accessible, started utilizing it to assist her write product requirement paperwork or PRDs. Each product supervisor is aware of what a PRD is. When Claire prompted ChatGPT to “write a PRD,” it didn’t want a protracted preamble. That one acronym carried a long time {of professional} observe. However Claire went additional. She refined her prompts, improved them, and taught ChatGPT methods to suppose like her. Over time, she had educated a system, not on the mannequin stage, however on the stage of context and workflow.
Subsequent, Claire turned her workflow right into a product. That product is a software program interface that wraps up various associated magic phrases right into a helpful package deal. It controls entry to her custom-made magic spell, so to talk. Claire added detailed prompts, integrations with different instruments, entry management, and an entire lot of conventional programming in a next-generation software that makes use of a mixture of conventional software program code and “magical” fuzzy perform calls to an LLM. ChatPRD even interviews customers to study extra about their objectives, customizing the appliance for every group and use case.
Claire’s quickstart information to ChatPRD is a good instance of what a magic-word (fuzzy perform name) software seems like.
You may as well see how magic phrases are crafted into magic spells and the way these spells are even a part of the structure of purposes like Claude Code by way of the explorations of builders like Jesse Vincent and Simon Willison.
In “How I’m Utilizing Coding Brokers in September, 2025,” Jesse first describes how his claude.md file gives a base immediate that “encodes a bunch of course of documentation and guidelines that do a reasonably good job maintaining Claude on monitor.” After which his workflow calls on a bunch of specialised prompts he has created (i.e., “spells” that give clearer and extra personalised that means to particular magic phrases) like “brainstorm,” “plan,” “architect,” “implement,” “debug,” and so forth. Observe how inside these prompts, he might use further magic phrases like DRY, YAGNI, and TDD, which seek advice from particular programming methodologies. For instance, right here’s his planning immediate (boldface mine):
Nice. I want your assist to write down out a complete implementation plan.
Assume that the engineer has zero context for our codebase and questionablestyle. doc every little thing they should know. which recordsdata to the touch for everyprocess, code, testing, docs they could must test. methods to check it.givethem the entire plan as bite-sized duties. DRY. YAGNI. TDD. frequent commits.
Assume they're a talented developer, however know virtually nothing about ourtoolset or drawback area. assume they do not know good check designverynicely.
please write out this plan, in full element, into docs/plans/
However Jesse didn’t cease there. He constructed a mission known as Superpowers, which makes use of Claude’s not too long ago introduced plug-in structure to “give Claude Code superpowers with a complete expertise library of confirmed methods, patterns, and instruments.” Asserting the mission, he wrote:
Abilities are what give your brokers Superpowers. The primary time they actually popped up on my radar was just a few weeks in the past when Anthropic rolled out improved Workplace doc creation. When the function rolled out, I went poking round a bit – I requested Claude to inform me all about its new expertise. And it was solely too completely satisfied to dish…. [Be sure to follow this link! – TOR]
One of many first expertise I taught Superpowers was Easy methods to create expertise. That has meant that once I wished to do one thing like add git worktree workflows to Superpowers, it was a matter of describing how I wished the workflows to go…after which Claude put the items collectively and added a pair notes to the prevailing expertise that wanted to clue future-Claude into utilizing worktrees.
After studying Jesse’s submit, Simon Willison did a bit extra digging into the unique doc dealing with expertise that Claude had introduced and that had sparked Jesse’s brainstorm. He famous:
Abilities are extra than simply prompts although: the repository additionally contains dozens of pre-written Python scripts for performing frequent operations.
pdf/scripts/fill_fillable_fields.py for instance is a customized CLI software that makes use of pypdf to search out after which fill in a bunch of PDF kind fields, specified as JSON, then render out the ensuing mixed PDF.
It is a actually subtle set of instruments for doc manipulation, and I really like that Anthropic have made these seen—presumably intentionally—to customers of Claude who know methods to ask for them.
You possibly can see what’s taking place right here. Magic phrases are being enhanced and given a extra rigorous definition, and new ones are being added to what, in fantasy tales, they name a “grimoire,” or e-book of spells. Microsoft calls such spells “metacognitive recipes,” an exquisite time period that ought to get broadly adopted, although on this article I’m going to stay with my fanciful analogy to magic.
At O’Reilly, we’re working with a really totally different set of magic phrases. For instance, we’re constructing a system for exactly focused competency-based studying, by way of which our clients can skip what they already know, grasp what they want, and show what they’ve discovered. It additionally offers company studying system managers the flexibility to assign studying objectives and to measure the ROI on their funding.
It seems that there are dozens of studying frameworks (and that’s itself a magic phrase). Within the design of our personal specialised studying framework, we’re invoking Bloom’s taxonomy, SFIA, and the Dreyfus Mannequin of Ability Acquisition. However when a buyer says, “We love your method, however we use LTEM,” we are able to invoke that framework as an alternative. Each company buyer additionally has its personal specialised tech stack. So we’re exploring methods to use magic phrases to let no matter we construct adapt dynamically not solely to our finish customers’ studying wants however to the tech stack and to the educational framework that already exists at every firm.
That will be a nightmare if we needed to help dozens of various studying frameworks utilizing conventional processes. However the issue appears rather more tractable if we’re capable of invoke the precise magic phrases. That’s what I imply once I say that magic phrases are an important constructing block within the subsequent era of software programming.
The Structure of Magic
Right here’s the essential factor: Magic isn’t arbitrary. In each mythic custom, it has construction, self-discipline, and price. The magician’s energy will depend on understanding the precise phrases, pronounced in the precise approach, with the precise intent.
The identical is true for AI methods. The effectiveness of our magic phrases will depend on context, grounding, and suggestions loops that give the mannequin dependable details about the world.
That’s why I discover the rising ecosystem of AI purposes so fascinating. It’s about offering the precise context to the mannequin. It’s about defining vocabularies, workflows, and roles that expose and make sense of the mannequin’s skills. It’s about turning implicit cultural data into express methods of interplay.
We’re solely firstly. However simply as early programmers discovered to construct structured software program with out spelling out precise machine directions, as we speak’s AI practitioners are studying to construct structured reasoning methods out of fuzzy language patterns.
Magic phrases aren’t only a poetic picture. They’re the syntax of a brand new form of computing. As folks turn out to be extra snug with LLMs, they’ll go across the magic phrases they’ve discovered as energy consumer methods. In the meantime, builders will wrap extra superior capabilities round present magic phrases and maybe even train the fashions new ones that haven’t but had the time to accrete enough that means by way of vast utilization within the coaching set. Every software can be constructed round a shared vocabulary that encodes its area data. Again in 2022, Mike Loukides known as these methods “formal casual languages.” That’s, they’re spoken in human language, however do higher while you apply a little bit of rigor.
And not less than for the foreseeable future, builders will write “shims” between the magic phrases that management the LLMs and the extra conventional programming instruments and methods that interface with present methods, a lot as Claire did with ChatPRD. However finally we’ll see true AI to AI communication.
Magic phrases and the spells constructed round them are solely the start. As soon as folks begin utilizing them in frequent, they turn out to be protocols. They outline how people and AI methods cooperate, and the way AI methods cooperate with one another.
We will already see this taking place. Frameworks like LangChain or the Mannequin Context Protocol (MCP) formalize how context and instruments are shared. Groups construct agentic workflows that rely on a standard vocabulary of intent. What’s an MCP server, in any case, however a mapping of a fuzzy perform name right into a set of predictable instruments and companies accessible at a given endpoint?
In different phrases, what was as soon as a set of magic spells is turning into infrastructure. When sufficient folks use the identical magic phrases, they cease being magic and begin being requirements—the constructing blocks for the subsequent era of software program.
We will already see this development with MCP. There are three distinct sorts of MCP servers. Some, like Playwright MCP, are designed to make it simpler for AIs to interface with purposes initially designed for interactive human use. Others, just like the GitHub MCP Server, are designed to make it simpler for AIs to interface with present APIs, that’s, with interfaces initially designed to be known as by conventional applications. However some are designed as a frontend for a real AI-to-AI dialog. Different protocols, like A2A, are already optimized for this third use case.
However in every case, an MCP server is mostly a dictionary (or in magic phrases, a spellbook) that explains the magic phrases that it understands and methods to invoke them. As Jesse Vincent put it to me after studying a draft of this piece:
The half that feels probably the most like magic spells is the half that the majority MCP authors do extremely poorly. Every software has a “description” area that tells the LLM how you utilize the software. That description area is learn and internalized by the LLM and modifications the way it behaves. Anthropic are notably good at software descriptions and most everyone else, in my expertise, is…much less good.
In some ways, publishing the prompts, software descriptions, context, and expertise that add performance to LLMs could also be a extra essential frontier of open supply AI than open weights. It’s essential that we deal with our enhancements to magic phrases not as proprietary secrets and techniques however as shared cultural artifacts. The extra open and participatory our vocabularies are, the extra inclusive and inventive the ensuing ecosystem can be.
Footnotes
- Whereas typically related as we speak with stage magic and cartoons, this magic phrase was apparently used from Roman occasions as a therapeutic spell. One proposed etymology means that it comes from the Aramaic for “I create as I communicate.”
