Let Your AI Invent the Tag — Then Match It to Reality

🚀 Quick Take

Tagging old content is boring, and if you've tried to automate it you've hit a wall: your vocabulary is too big to stuff into a prompt. Blog author Simon Willison ran straight into this. His site carries 1,856 tags — too many to feed an LLM and ask it to pick the right ones.

Instead of forcing the model to classify against a giant list, Simon Willison highlights a trick from Doug Turnbull: let the model invent brand-new tags with zero knowledge of your existing vocabulary, then use vector embeddings to find which real tags sit closest to what the model imagined. You halve the problem. The model free-writes, the math does the matching.

The lesson scales way past blogging — it's a pattern for any workflow where you have a fixed taxonomy but an unruly pile of new content. For crypto researchers, that's every day of the week.

🛠 What It Is

The approach is simple and worth unpacking because it's counterintuitive. In a classic classification setup you hand the LLM a list of candidate categories and ask it to sort the input. That breaks when the list is enormous. Not because the model is dumb, but because a 1,856-item tag list is a token furnace and a recipe for inconsistent output.

The fix inverts the pipeline. You ask the model to generate fresh tags from scratch, guided only by a sample of the shape you want — the source example shows a furniture store feeding the model a snippet of its real hierarchy, like "Furniture / Living Room Furniture / Coffee Tables & End Tables," so the inventiveness lands in the right shape.

Then you embed both the hallucinated tags and your real tag vocabulary into vectors, and find whichever real tags are closest to the imagined ones. The model suggests; similarity search confirms. The output is a tag from your actual vocabulary, perfectly matched.

The source example is retail, but the pattern is agnostic. Any domain with a structured taxonomy — support ticket categories, compliance rules, token listings — can use it.

🧠 Why Traders Should Care

This is bigger than tag hygiene. The pattern is a clean answer to a problem every trader faces constantly: too many candidates, too little time, and a mental model of what's "good" that's too big to articulate.

You don't need to be a developer to benefit. If you're building any kind of screen — an alert filter, a watchlist, a research folder — you already have a personal taxonomy in your head. The problem is you can't type it all out, and you definitely can't type it out fresh every time a new token drops. This pattern teaches a smarter workflow: sketch the shape of what you're looking for, let the tool generate candidates, then anchor them against known data.

There's a deeper point for anyone trading memecoins and low-cap tokens. Classification against a fixed list is how you miss things. New projects are, by definition, inventing categories that didn't exist before. The winning filter isn't the one that sorts accurately against yesterday's categories — it's the one that can hypothesize a new category and then validate it. That's a very different muscle, and it's the one this technique exercises.

The other takeaway: most of the drudgery in this game is matching. You find a token, then you have to figure out where it sits, who the holders are, whether the liquidity is locked. Matching every alert against a full security checklist manually is exactly the kind of 1,856-item-tag problem that should be automated.

⚡ Put It To Work Today

You don't need to wait for a tag-cleaning weekend. Start small.

If you have any recurring analysis task — a weekly review of projects you've tracked, a research notes file, a Telegram archive you want to search — apply the two-step pattern. Step one: draft freely, no constraints. Step two: match the draft against your real records using similarity. You'll get results that fit your actual system instead of generic-sounding categories.

For the crypto side specifically, start with a tool that's already done this matching work for you. The Blackhat Empire free alert network on Telegram — entry point @gmgnalerts — runs 450+ groups with live buy and sell alerts across SOL, BSC, ROBINHOOD and more. Every alert already passes through a layered security gate covering GoPlus, RugCheck, GMGN entrapment and bundler checks, holder analysis, and LP lock-burn verification. The red flags are printed directly on the alert. That's your classification step, handled.

Then XTRACK (@xtrack1bot) auto-tracks every alerted token and pings you at each multiplier milestone, attaching holders, LP status, and security data as it goes. @VBMBbot handles multibuy scanning. The web terminal at blackhat.finance gives you trending lists and the DYOR Academy articles. The alerts deep-link straight into GMGN, which you can register free via gmgn.ai/?ref=10Xboost.

Try one workflow this week: let an AI draft a short research note on a token without showing it your usual criteria, then compare where it landed versus where you'd have put it. You'll find the gaps in your own filter.

🎯 Bottom Line

The "hallucinate first, match second" pattern is a genuinely useful reframe. It acknowledges that your internal vocabulary is too rich to dump into a prompt, and that your time is better spent validating good guesses than generating them from scratch.

Whether you're tagging a decade of blog posts or scanning a token stream, the same principle holds: don't force the big list into the prompt. Let the tool invent, then measure the distance between invention and reality.

Credit where due — this walked out of a 14 August 2026 post by Simon Willison, riffing on Doug Turnbull's concrete example. The technique deserves a wider audience than the blog-tagging niche it came from.

Join the free alerts: @gmgnalerts on Telegram — every signal pre-screened, every red flag printed before you click.

DYOR: This is informational, not financial advice. Crypto is risky; verify every contract and security signal yourself.

🏴 Blackhat Empire — Free Alert Network

🚪 Telegram Portal: @gmgnalerts 📲 Trade on GMGN: gmgn.ai 📍 Live plays & full DYOR: blackhat.finance 🏴 Add all 7 MAIN groups: t.me/addlist 💬 Community Chat: @gmgnx_chat 🤖 Power tools: @VBMBbot · @xtrack1bot