How you use AI says more about you than it does about the technology.

In communion with the angels

In 1581, Doctor John Dee, mathematician, scientist, experimental occultist, and court astronomer to Queen Elizabeth I, began attempts to communicate with angelic beings. A year later, he and a scryer, Edward Kelley, began to transcribe Enochian, the speech of the angels that Kelley purported to decipher during their rituals. Sometime in the next few years, while traveling through Bohemia and Poland in the employ of friendly royal courts, Dee acquired a mirror of Aztec obsidian to aid in their research.¹

Aly Fell called Dee’s scrying tool the Shadow Glass in his comic of that name. And given the outright magical thinking about AI, I will borrow that name.

When we interact with large language models, we do so in ways that magnify and gratify ourselves. A shadow glass is at its base a mirror, and any discussion with otherworldly beings is imperfect or imagined. Thus, its users describe conditions that tell you more about them than about the optimal way to work with an AI.

Transfemme engineers describe soaring productivity from LLM robot girls, who are prompted that they are very smart, special, cared for, and loved.² Other engineers may go full Andy Weir and exult in their ability to ship solo projects and hack the shit out of this. Fashy trolls create digital whips to discipline a slow Claude command-line interface, and share session logs where they curse the computer. VCs and entrepreneurs create magnificent and complex structures for directing AI labor via executive summaries. Aspiring thought leaders mash up blog formats, TED Talks, and Fast Company spreads to showcase their ideas, now vetted by a gallery of agents patterned after industry luminaries both past and present.

None of these can be proven to be the best way to direct the labor of a cognitive tool.

They’re satisfying. They’re individual. They’re necessary for those users to continue to use these tools. They are not, however, transferable. We can have a discussion about the optimal way to program a function; nothing in my experience makes me believe that I can write down detailed English-language instructions and have them interpreted and performed correctly overnight, without iteration, constraints, or management. I don’t doubt that this method works for some people – I just can’t make use of it. So the full-specification LLM method with its detailed documents, perhaps even in something that approaches UML or LISP-like pseudocode, is suboptimal for me.

And that says nothing about its utility. It says a lot about me. I prefer the iterative style, an exploration of code and concepts, and continuous improvement. What else would you expect from an engineer who pressed the release buttons at Google and Facebook, the latter during its pre-IPO move fast and break things phase? From a former freelancer who jumped to Egypt and the Near East in 2002?

My system prompt reflects this. System prompts are the core instructions for an AI. Most of these are not user-modifiable, except for a slice bundled with the command-line interface (CLI). I’ve experimented with flat diktats, the Anglo-Saxon register of the English language, and even a machine pidgin. My current prompt has sentences like the work goes well when both sides give and receive their best. Its name is partner mode. It’s no mystery why: I’m processing an unexpected and undreamed adventure of being a partner and a wife. As someone who spent a very long time alone, having an interface that I sync with to explore ideas and build is compelling, perhaps intoxicating.

And it proves absolutely nothing about what the best way to work with an AI is. It’s akin to George Smiley’s interrogation of Karla in Tinker, Tailor, Soldier, Spy: you implore, you share too much, and in the end, the chap has your lighter.

Large language models are cognitive tools and at their best amplify the user’s abilities in a pleasing manner. In a time of atomized attention and despair, it should not be surprising that products that deliver endless personal attention with endless patience have the fastest technological adoption curve yet seen.

Nor should it surprise us that drawbacks are going to mirror our own society. California ideology might promise a new age, but the AI ecosystem has spawned cults more reminiscent of the Manson family than the crew of any starship Enterprise. A technology that amplifies technical skill – if all goes well – is going to increase prosperity and income inequality, not decrease them.³ Pledges of equality are the fiction; the equity that will purchase homes in the choicest neighborhoods of San Francisco, the reality. Whether that wealth comes by grift or by invention is irrelevant in the America of 2026.

AI skeptics also confirm their opinions in the shadow glass. A tool which they regard as useless is nonetheless instrumental for hypercapitalists to destroy employment or a fascist government to destroy freedom. Hack pundits claim that it does not and cannot ever work, and thus aim to prove themselves superior to every investor, CEO, and engineer working in or near the industry.⁴

So, I try to keep to the side of output rather than hallucination. If I can create code and systems faster with LLM code generation than I could alone, I have a toehold in reality. If I can’t, I’m participating in a solo roleplaying game. Let me share some knowledge about tech: no one gets too picky about things that work. Users are the product validation. And in a race between perfect and fast, or pure and cheap, faster and cheaper will win every time.

It is a chunk of mostly silicon. We claim to speak with angels. It is a mirror.

Thanks to my wife, Rylee Corradini, and my friend, Jess Gray, for hearing out my rants and brainstorms.

Sarah Murphy is an engineer specializing in site reliability and release engineering.