Loooom: Curated Skills for People Who Don't Code

I built Loooom as a personal experiment. The goal was never to win a market. It was to learn two things I couldn't find good answers to anywhere else: how skill marketplaces actually work, and how you tell a good skill from a bad one.

After a few months of building, the second question turned out to be the interesting one.

What I tried first

The first version of Loooom was a full social platform — SvelteKit, Postgres, auth, content-addressed versioning. Then Anthropic shipped the plugin marketplace spec and I realized I was rebuilding GitHub, so I pivoted to a GitHub-native marketplace: a repo, a marketplace.json, install with /plugin marketplace add mager/loooom. Simpler, and it worked.

Along the way I also bolted on two side experiments: ME.md (a portable file describing a human for an agent to read) and an "agent soul" format. Three separate ideas, three separate things that each needed their own audience.

None of it got much use. Two things were actually wrong:

1. I was building a marketplace with no network. Marketplaces have a brutal bootstrapping problem — no supply attracts no demand, which attracts no supply. Solving that takes either a community you already have or a lot of luck. I had neither, and three half-built protocols competing for the same attention didn't help.

2. The audience was wrong for the format. A Claude Code plugin only reaches people who run Claude Code. That's developers. But the most interesting skills — the ones with real expertise behind them — usually belong to people who will never touch a terminal. The format I'd chosen excluded exactly the people I most wanted to publish.

The part that kept being interesting

What held my attention the whole time was quality. Most "skills" you find in the wild are a paragraph of generic instructions — a prompt with a nice filename. They don't carry any actual expertise. You could regenerate them from the title alone.

A small number are genuinely good, and the difference is specific and learnable. A good skill has a real method, not vibes. It has concrete examples. It has a point of view — it tells the model what not to do, which is usually where the expertise lives. When I started writing evals to score skills (I've written about that before), the good ones held up under scrutiny and the rest fell apart.

Recently I made that concrete. Every skill now gets scored against a written rubric — by a model, not by me — and because I'm not paying for an experiment, the judge runs on a free model. The check that matters is whether it can fail a bad skill. A deliberately generic version of one of my skills scores an 8 out of 100; the real one scores 100. If the judge couldn't tell them apart the scores would be theater, so that separation is the thing I actually test for.

The other thing I kept noticing: almost nothing I use Claude for is about code. I've leaned on it to learn Japanese before a trip, plan meals, and rewrite emails I was dreading. The terminal is incidental — the value is the expertise on the other end. If that's true for me, the skills worth collecting are mostly the non-technical ones.

That gap — between a skill that's a costume and a skill that's a craft — is the thing I actually want to study. So I'm pointing Loooom directly at it.

The new shape

Loooom is becoming a curated collection of non-technical skills, held to a quality bar I set deliberately high. Not an open marketplace. Closer in spirit to impeccable.design, which isn't a UGC catalog at all — it's one tight, well-made suite of design skills. The cohesion is the product. The taste is the product.

The supply problem disappears when you stop pretending to be a marketplace and just author the things yourself. So that's what I did. The first set, Volume 1, is fifteen skills for the parts of life that have nothing to do with code: cooking, writing, money, negotiation, storytelling, public speaking, focus.

They share a house style borrowed from impeccable: one-word names. hook, voice, story, improvise, leverage, stage, stack, decide. A grab-bag of topics reads like a collection when the naming is consistent.

All fifteen are written and live now, in four clusters: creativity (hook, voice, story, frame), everyday life (improvise, train, rest, stack), communication (leverage, stage, reply, level), and the mind (decide, focus, absorb). I started in my own wheelhouse — music and writing — where I could judge the result myself, then worked outward into things I'm less expert in, leaning on the rubric to hold the bar where my own taste ran out. Every one of them clears it.

Here's the difference I'm chasing, concretely. A generic version of hook would say "write a catchy chorus and use repetition." The actual skill says: find the one line the whole song is about; put the title where the ear expects it, which is the first or last line of the chorus, because the ear remembers edges; put open vowels on the high notes because tight vowels choke; repeat the melodic shape but change exactly one element so it's familiar but not boring. Then it works through a real example end to end, and it tells the model to make you say the song's idea in one sentence before writing anything.

The test I'm holding every skill to: would someone who actually does this for a living recognize it as real?

Tradeoffs I'm making on purpose

Curated, not open — for now. Open submissions are the eventual goal, but you earn that by building something good enough that people want in. Quality first, scale later.

ME.md and the agent-soul format are parked. They're interesting, but they're separate flywheels, and the lesson of the first few months was to stop running three experiments at once.

What stays the same

Skills are free, and always will be. The whole thing is open source. And the name still means what it meant: a loom is where threads become fabric. Loooom — four o's, like extra yarn on the spool — is where knowledge becomes something an agent can actually use. My wife knits; I like that this feels more like a yarn shop than a tech company.

What's next

1. Tighten the rubric and the judge. Fourteen 100s and a 97 is a little too clean — a sharper judge should spread the scores out more, and I want to trust them before I lean on them harder
2. Start Volume 2, and keep pushing past my own expertise
3. Eventually: open submissions

If you've read this far, the honest summary is that Loooom is a lab, not a startup. The product is the experiment, and right now the experiment is: how good can a skill actually be, and what does it take to keep a collection of them at that level?

---

Loooom is open source and always will be. Skills are free and always will be.

Built in Chicago by @mager and @magerbot.