Sarah Murphy has written something that's been bouncing around my head since I read it. In a post called "The Shadow Glass," she takes a 16th-century occult artifact, an obsidian mirror used by John Dee to commune with angels, and uses it as a metaphor for how we talk to LLMs. Her argument is simple and sharp: the way you use AI tools tells us more about you than about the tool.
She catalogs the archetypes she's seen. Transfemme engineers who prompt their AI assistants with affirmations about being smart and loved. Solo hackers shipping projects at a furious clip. VCs building elaborate command structures. None of these is the right way, and good taste is often what separates the effective users from the rest. They're personal rituals that work for specific people, and they're not transferable.
Murphy knows this because she's tried the methods that work for others and they don't work for her. She's an engineer who shipped releases at Google and Facebook in their move-fast eras. Her system prompt, called "partner mode," emphasizes mutual exchange. It includes lines like "the work goes well when both sides give and receive their best." Her approach is personal, not prescriptive, standing in contrast to the cognitive surrender that often accompanies AI reliance. The piece goes further than personal quirks. Murphy argues that AI adoption is driven partly by what the tools provide emotionally: endless patient attention in a time of "atomized attention and despair." And she's blunt about the economics. A technology that amplifies technical skill will increase income inequality, not reduce it. The California ideology promises a new age, but the wealth ends up buying houses in San Francisco's best neighborhoods, whether through grift or invention. She's also direct about skeptics, who she says are just staring into the same mirror, confirming their own biases about a tool they claim is useless. A Hacker News commenter pushed Murphy's thesis further, arguing that the mirror quality is the feature, not a bug. LLMs are best used to extract ideas from your own mind, the stuff that sits "below verbal mind," rather than just directing them outward. That reading reinforces the core insight. The most productive AI relationships come from understanding your own cognitive patterns, not from finding some universal prompt formula. It's a chunk of silicon. We claim to speak with angels. It's a mirror.