--- title: "The Start of Content" order: 30 status: draft visible: true --- ## First Words I didn’t start with “mythic.” That came later, and honestly more as a label than a goal. At the beginning, the only real anchor was Tarot. I knew I wanted those cards in the deck—they carry a kind of built-in structure and recognition that would be hard to recreate from scratch—so the question became less “what words do I like” and more “what kinds of words can exist alongside Tarot without feeling out of place.” That led to a lot of early probing in different directions. I tried thinking about parts of speech first, just to get some footing. Nouns settled in pretty quickly as the most natural fit. Verbs felt too directive, like they were telling you what to do instead of representing something. Adjectives didn’t really hold up either—they felt thin, more like modifiers than anchors. Tarot cards don’t describe something, they *are* something, and that pushed things pretty firmly toward nouns as the core. From there, tone became the next pressure, although I wouldn’t have called it that at the time. It was more of a repeated question: does this feel like it belongs here? I tried more literal words, which tended to feel flat, and more abstract ones, which often drifted too far into vagueness. There’s a middle space where a word has some weight to it, but still leaves room for interpretation, and it took a while to even recognize that space, let alone operate inside it consistently. At the same time, I was already trying to layer in constraints that had nothing to do with tone at all. One of the big ones, taken directly from EverDeck, was the idea that the words shouldn’t just be labels—they should carry some additional meaning or utility. In this case, that meant paying attention to first-letter distribution so that the deck could support letter-based mechanics. That immediately complicates things, because now a word isn’t just being evaluated on meaning or feel, but also on whether it fits into a larger distribution that has to balance across the entire deck. That constraint alone caused a surprising amount of churn. You might find a word that fits perfectly in tone, but it lands on the wrong letter. Or it fits the letter, but now you’ve overloaded that letter category. So even early on, before anything else was really in place, there was already this tension between “this is a good word” and “this is a word that actually fits the system.” There were other filters as well. No duplicates, obviously, but also a general effort to avoid words that were too close semantically. That wasn’t perfectly enforced—there are definitely a few cases in the final deck where things overlap more than I’d like—but it was a consistent pressure during selection. If two words felt like they were doing the same job, one of them probably didn’t belong. And then there was the visual side, even at this early stage. I wasn’t drawing anything yet, but I had to at least believe that a word could be represented simply and clearly. Some candidates fell out immediately once I tried to picture them. Either they required too much detail, or they were too abstract to ground in a scene without getting convoluted. That constraint was softer than the others, but it still eliminated a fair number of otherwise decent options. The AI process during this phase was… uneven. Early attempts were more about trying to get it to produce something close to a finished list, balancing all of these constraints at once. That didn’t really work. The outputs would look reasonable on the surface, but drift under scrutiny—clusters forming where they shouldn’t, tone inconsistencies, letter imbalances creeping in. Tightening the prompts helped in some ways, but often made the results feel overly uniform, like it was optimizing for the prompt instead of the actual problem. The shift, eventually, was to stop expecting a solution and instead use AI to generate material. Large pools of candidate words, suggestions for specific gaps, alternatives for weaker entries. Less “solve the deck” and more “help me explore this space.” Around that time I moved more of the process into paid ChatGPT, mostly because it handled longer context and iteration better than the earlier tools I had been experimenting with. It still didn’t solve the problem, but it was much better at staying within the boundaries I was trying to establish and helping refine specific areas without everything drifting. Over repeated passes—generate, review, replace, rebalance, repeat—the word set started to stabilize. Not in the sense that everything was perfect, but in the sense that nothing was obviously out of place anymore, and the constraints were being satisfied without constant breakage elsewhere. Somewhere in that process, a pattern became more visible. The words that tended to survive weren’t necessarily complex. In fact, a lot of them were fairly simple on the surface. But they carried a kind of weight in how broadly they could be interpreted, or how easily they could map to different contexts. They could mean slightly different things depending on how they were used, which made them more flexible without being vague. *Seed* is a good example. It’s easy to picture, and it can also be the beginning of an idea or something that hasn’t happened yet. A *Rod* can be something that is passed, a symbol of violence, or a sign of authority. *Deluge* and *Abyss* have that same foothold in something physical, but they carry a mental or emotional tone with them too. And then there’s *X-roads*, my one cheat for the letter X. I bent the spelling, but I liked what the word gave me: a place you can picture and a choice you have to make. That’s what eventually got labeled as “mythic,” although that label came after the fact. It wasn’t something I set out to achieve so much as something I noticed once enough of the right kinds of words had accumulated. By that point, the word list wasn’t really a loose collection anymore. It was starting to feel like a constrained vocabulary—something that could support structure, even if that structure wasn’t fully defined yet. And that’s where the next layer starts to push back on it.