Where this started
I love SREF codes. I kept losing them. So I built the bookmark manager I wished existed.
I’d been collecting MidJourney style references the way most heavy users do — bookmarked X posts, saved YouTube links, screenshots in a folder, the occasional Notion table — and it failed me every time I needed to find one back. The good curation sites and channels existed, but retrieval across them was a mess. At some point the friction got loud enough that I noticed it could be a thing I built. So I did, and I’ve been using it nightly ever since.
That’s the whole motivation. The mother of the initiative was friction in my own process.
What an SREF actually is
SREF codes are MidJourney’s stylistic shortcuts — appended to a prompt as --sref <number>, they pull in a discoverable subset of MJ’s style training and lock a look in without you having to describe it.
If you don’t live in MidJourney, the short version: SREFs are numeric tags — one number per discoverable visual style — and once you find one that hits, your prompts get to focus on subject, framing, composition, and context, while the style holds with near-perfect fidelity. You can also chain them (a “suffix” can carry multiple SREFs that blend), which means a good SREF stack is its own kind of authored asset, worth saving. SMC saves them: up to four example images per code, a title, description, and filtering tags. Search by text, browse the library visually, filter by tag, copy the suffix to your clipboard. That’s the loop.
The build
~300 hours, August through December 2025, in the 10pm-to-midnight window most new fathers spend asleep — my first end-to-end agentic ship.
I built SMC in the bookend between bedtime and burnout, with Subframe handling the design-system layer, Claude Code doing the heavy lifting in the codebase, and Cursor as my editor — mostly for its browser-integrated agents and, honestly, for its commit-message generator, which I’d pay the whole subscription for on its own. Supabase under everything.
Calling 300 hours the “build” is generous. If you net out the false starts and the rabbit holes, the version that’s live today is more like 80 hours of work. The other 220 was the education. Worth every one — but I’d want to be clear-eyed about which is which.
The Chrome extension that almost got everyone banned
A coworker suggested a companion extension to capture codes faster from inside MidJourney. I spent a month on it. Then I asked Manus, and Manus said stop.
The pitch was good: a Chrome extension that adorned the MJ UI for any user logged into both apps, one-click capturing a code with its preview image straight into their SMC library. I’d done the up-front market-viability pass with Manus before development and hadn’t touched it in months. A few weeks into the extension build, mostly to be safe, I went back and ran the idea past it.
Manus stopped me cold. Adorning the MidJourney UI inside a user’s logged-in session would put me — and every one of my users — on the wrong side of MJ’s terms of service. Even if the extension worked, the people I built it for could lose their accounts. That ended the extension.
What we landed on instead is the workflow that’s in the product today: open MidJourney in one browser window, SMC in another, drag any image from MJ onto SMC, and it kicks off a save with that image already attached. Drop up to three more related images into the dialog and you’ve got a four-image card with the original code reference still pulled cleanly off the MJ side. Friction down hard, EULA intact. Better than the version that would have gotten people banned.
Three starts
I rebuilt the project from scratch twice before the third version stuck. Supabase was the only thing that survived all three.
I started in MagicPath, abandoned it, spent a couple of nights in OnLook — a then-very-new Figma alternative that looked promising on the surface but that I couldn’t get my head around inside the time I had — and finally landed in Subframe, which is what got me from design system to shipped UI. If I were doing it again today, I’d skip Subframe: Figma to clickable prototype in Claude Code, unify the design language inside the prototype, then straight to production. Subframe got me over the hump, but its one-way workflow — design system upstream, codebase downstream, wrapper components in the repo so syncs don’t clobber your edits — was painful when the designer and the developer are the same person. I still like Subframe; a tight Subframe / Claude Code skill might pull me back. As of right now, I’d start somewhere else.
Code is cheap, and learning when to drop what isn’t serving you is the more useful muscle. Starting over is not a dirty word anymore — at least not for me.
About the name
SMC stands for SREF Mining Company. The mining is literal — at launch, the best way to find a new SREF was to type --sref random and let MidJourney pan you a number out of around ten billion.
--sref random returns one number from a very large numeric range (around 10¹⁰ at launch, growing — there are three numeric sets now where there were two then, and the UI is still catching up). Each number is a potential visual style, most are unremarkable, a few are gold. Finding the ones worth keeping looked a lot like panning. The free tier is the crowdsource side of that — anyone can browse, anyone can sign up, and the public library is the shared assay table.
Where it stands
Live, free, and used by approximately three people.
The public library is open to anyone — browse and search without an account. The signup gate kicks in only when you want to copy a code’s suffix to your clipboard or save your own. From there: personal library (still public on the free tier), favorites from the community library, the works.
User count, candidly, is roughly three: me, my wife, and a couple of coworkers who humored me. I didn’t plan the marketing side, and it shows. What I have done is leave a standing offer — anyone visibly active in the SREF community who signs up gets a paid-tier account, free, in perpetuity, even after I eventually flip the paid tier on. I’d planned to give away fifty of those and I’d extend it past a hundred if there’s any sign of traction and the costs stay reasonable. So far no notable SREF curator has taken me up on it. Offer’s still open.
What this was actually about
Proof of build, not proof of market. The lesson was that I could do this — fast, completely, between 10pm and midnight, with an infant in the next room.
SMC is the first thing I’ve shipped end-to-end on an agentic stack, and the most useful thing it taught me is the shape of my own build process — what I can do in narrow windows, where I waste cycles, when to throw work away. The constraint of two hours a night with maybe seventy-five minutes of usable focus inside it forced a discipline of “good enough” decisions I wouldn’t have arrived at with eight hours and the luxury to wander. If I built it again with daytime hours and what I know now, my honest guess is 40–50 hours.
It’s a useful little app, it’s the one I use most personally, and it’s still the cleanest evidence I’ve got that the loop works — that one person, a design tool, an agent, and a backend can take an idea to shipped product without anyone else in the loop. The market piece I’ll worry about when there’s a reason to. The build piece is in the bag.