Verses Over Variables: How To

The Ghost Prompts in My Folders

At some point in late 2023, I started naming my prompt files like they were doomed expeditions. "linkedin_hook_FINAL_v7_USE_THIS_ONE" lived next to "brand_voice_ACTUALLY_GOOD_this_time" lived next to a document called "prompt_graveyard_DO_NOT_DELETE" which I absolutely should have deleted but couldn't because what if I needed to resurrect that one paragraph about strategic frameworks that took forty-five minutes to perfect and then stopped working three weeks later for reasons I still don't understand. This is not a system. This is a haunted library where every book is about yourself and all of them are slightly wrong.

The promise of AI was supposed to be that you'd explain what you wanted once, the machine would learn, and then you'd both move on with your lives like functional collaborators. Instead we got a technology that suffers from the worst case of short-term memory loss since Guy Pearce in Memento, except the tattoos keep fading and you have to keep re-inking them because the context window filled up again and now Claude doesn't remember that you prefer sentence fragments for emphasis or that you're allergic to the phrase "leverage synergies" or that time you spent an hour explaining your brand voice which the AI absolutely understood and then completely forgot by Tuesday.

And then in October 2025, Anthropic shipped something called Skills that changed the fundamental architecture. You can now teach Claude something once and have it stick. Permanently. Across every conversation. Forever.

Skills are folders containing Markdown files with instructions. That's it. That's the whole technical implementation. A skill.md file with some metadata telling Claude when to load it, followed by whatever expertise you want to codify. The system scans available Skills using about 100 tokens per Skill to figure out relevance, then loads the full instructions only when they matter. This means you can maintain an entire team of specialists without overwhelming the context window, which is the technical problem that's been haunting every attempt to make AI systems remember things.

The boring technical format actually matters. Because Skills are just markdown files in a folder (occasionally zipped if you want to bundle resources), they're not locked to Claude. You can take that same skill.md file, drop it into a custom GPT, feed it to Gemini, hand it to whatever AI system launches next month that we're all supposed to be excited about. The expertise you spent thirty hours codifying doesn't evaporate when you switch platforms or when Anthropic decides to sunset a feature or when your company mandates a different vendor. This is the kind of mundane portability we used to assume before everything became a walled garden with proprietary formats that only work inside one company's ecosystem (and only until the next API breaking change). It's deeply unsexy infrastructure stuff that nobody will write breathless blog posts about. It also means you're building actual assets instead of renting capabilities from a vendor who might change the terms tomorrow. The ghosts stay with you, in other words. Even when you move. The metaphor that keeps coming back is ghosts. Not horror-movie ghosts. More like the feeling of walking through your childhood home after your parents have moved out. Everything's familiar and wrong simultaneously. Your old prompts are still there in that folder. They still contain your thinking. But they're artifacts from a different era of how this technology worked, and going back to them now feels like trying to have a conversation with a version of yourself who didn't know something fundamental about the system's actual capabilities.

Here's what actually changes because of Skills: You stop managing prompts and start managing specialists. Instead of asking Claude to be a brilliant copywriter AND a visionary art director AND a sharp strategist in the same breath (which is like asking your doctor to also fix your car and prepare your taxes), you build separate experts who do different things. The mechanics work through Claude acting as a manager that delegates automatically. You don't explicitly invoke Skills. You just describe what you're trying to do, and the system figures out which experts to pull in. The context-switching happens in the system instead of in your head. I built three specialists for content generation over about thirty hours of work, which feels like a lot until you realize it saves you hundreds of hours over the following months through not having to re-explain your creative voice every single day forever. The process involves gathering your best work, analyzing your style patterns, then codifying them into instructions. The first draft is always wrong. So you argue with it. You say "no, the thing about sentence fragments isn't that I use them a lot, it's that I use them for emphasis at specific emotional beats." And then you test and adjust until it works. (I realize I sound unhinged caring this much about workflow optimization for content creation. But we contain multitudes. Some of those multitudes are deeply invested in not wasting time on repetitive tasks. The rest are embarrassed about the first set. Both can be true.)

The thing that's genuinely different: Skills work across every conversation automatically without you managing them. Custom instructions apply universally but can't be specific. Memory systems try to learn from history but get confused. Projects isolate context but require you to remember which project you're in. Skills sit in the middle ground. Always available, only activate when relevant, persist across all your work. This represents a shift from "AI as magical black box" toward "AI as trainable infrastructure," which is less exciting but more useful. The magic black box approach leads to disappointment when reality doesn't match inflated expectations. The trainable infrastructure approach produces steady improvements because you're explicit about what you want and iterating toward it. Teaching beats hoping. Always has, probably always will.

Most people will read about Skills and think "sounds complicated" and continue suffering through the prompt library approach. The early adopters who invest the setup time will have a compounding advantage. Not because Skills are difficult but because building good ones requires domain expertise and taste. You need to know what excellence looks like in your field before you can codify it. That knowledge takes time to develop and can't be easily replicated.

Three years from now, someone will write about the moment AI systems became actually useful for creative work, and they'll probably point to Skills as the turning point. Not because the technology was revolutionary (it's just Markdown files) but because it solved the memory problem that was making everything else annoying. The ghost prompts will be historical artifacts. "Remember when we had to manually re-teach the AI our preferences every single day?" Future workers will find this as baffling as we find carbon paper. Obviously the system should remember. The fact that we tolerated the old way for two years says more about how quickly we adjust to frustrating patterns than about whether those patterns were actually okay.

I went back to that folder yesterday. The one with all the ghost prompts. They're still there, slightly dusty, occasionally useful for reference. Some of them are actually good. The writing is clear. The frameworks are solid. They just don't fit the shape of how the technology works anymore. The folder will probably stay on my hard drive forever, not because I need it but because deleting it would mean admitting something has actually changed. The lights are still on in the office. You can still see the old prompts if you look. Nobody's using them anymore. But the system remembers.

I have a folder of Skill templates I'm willing to share, which is ironic given the whole ghost prompts thing, but here we are. If you want to see the actual files I use, or if you need help building specialists for your own work, email me at [email protected]. The ghosts are friendly, I promise.

Share your results in the comments and let me know what worked.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

we’ve also started publishing more frequently on LinkedIn, and you can follow us here

if you’d like to chat further about opportunities or interest in AI, please feel free to reply.

if you have any feedback or want to engage with any of the topics discussed in Verses Over Variables, please feel free to reply to this email.

banner images created with Midjourney.