super Lewis

Ex-PM building a one-person company with AI. 4 products, real revenue. Sharing how product thinking turns solo-built tools into a profitable portfolio.

[super Lewis @superlewis88]

https://t.co/mPk4zv7fqv

[super Lewis @superlewis88]

https://t.co/LD5feStsOR

[super Lewis @superlewis88]

https://t.co/Px57LcZ9on

[super Lewis @superlewis88]

Still adding MiniMax H3 prompts. 224 now.

Every one sits next to the clip it actually made — hover to play. Some the authors published, some we reconstructed from the video. Labelled which is which.

https://t.co/sFhc6RymX5

Repo if you'd rather grep it: https://t.co/VZsUvSxKoM

[super Lewis @superlewis88]

@DasNripanka Honestly, it comes down to the user's own skill. Video generation still rewards technique right now — the people pulling clean results are the ones who know how to prompt and structure references properly. The model sets the ceiling; getting there is still on you.

[super Lewis @superlewis88]

MiniMax H3 shipped last week. We spent a few days going through what the community built with it. What stands out:

Native 2K with audio in the same pass. 2560×1440, 24fps, 4–15s. The audio isn't bolted on — ambience, foley and dialogue land on the frame you specify. Write "heels on tile," "a distant train," "she says quietly: you're late," and it's there.

Omni Reference — 9 images + 3 videos + 3 audio clips in one request. One creator drove an entire playable-looking game sequence with 8 images (6 characters, 1 arena, 1 UI), and H3 held every identity, the UI and the action logic coherent across the whole clip. This is where it pulls ahead of other video models.

It follows long prompts. Our longest is 6,602 characters: per-second shot breakdown, exact on-screen text, anatomy locks, focal lengths, a frame-accurate audio cue sheet. H3 hits the beats. You can write a shot list instead of rolling dice. Plus generative editing: first/last-frame control, motion transfer, single-element changes.

So we collected the prompts — 136, each shown next to the clip it produced. 64 are author-published (credited, linked back). 72 had no public prompt, so we reconstructed them from the footage and labelled them as reconstructions.
Straight up on those: we sample 8 frames, hand them to a vision model, and write the prompt that would most plausibly reproduce the clip. It describes the output — it can't recover negative constraints, exact dialogue, or reference-image workflows. A writing reference, not the author's prompt. Every card says so.

Repo's open too: our 72 reconstructions are MIT, full text. Author prompts are indexed, not copied. Wrote one and want it gone? Open an issue.

Gallery → https://t.co/AU0zlNJcWg
Repo → https://t.co/fNS4oxqZmI
https://t.co/iwiEaBUJgB runs 85+ image, video and language models behind one key — GPT Image 2, Nano Banana 2 & Pro, Seedance 2.0, Kling V3, VEO 3.1. H3 isn't on our API yet; we don't list models we haven't got working. When it runs, it ships — and this library updates the same day.

[super Lewis @superlewis88]

June's revenue was only $3,000 — it grew 2.2x in a single month.

[super Lewis @superlewis88]

$9660 MRR (July) — one person, 4 products, no team, no funding. I'm a product manager who learned to ship with AI. No CS background, no viral launch. Just picking the right problems and letting people find them through search. This pinned post is my scoreboard — I'll update the number every month and break down what moved it. If you're building solo, follow along.

[super Lewis @superlewis88]

@Preda2005 These character designs are fantastic. Would you be willing to share the framework behind your prompt? Just the structure, not the specifics — would really appreciate it!

[super Lewis @superlewis88]

Not seeing Opus 5 when you run /model?
You're on an old version of Claude Code.
Just update it — like this 👇 — and go see what Opus 5 can really do.

[super Lewis @superlewis88]

Opus 5 just dropped. In my own testing it beats the nerfed Fable 5 — and the price didn't move.

Every Anthropic model update makes me a little more confident, and a little more at ease, about building products.

What a time to be building.

[super Lewis @superlewis88]

https://t.co/t8TvrvK0Sb

[super Lewis @superlewis88]

$2997 MRR (June '26) — one person, 4 products, no team, no funding.

I'm a product manager who learned to ship with AI.
No CS background, no viral launch.
Just picking the right problems and letting people find them through search.
This pinned post is my scoreboard — I'll update the number every month and break down what moved it.
If you're building solo, follow along. 👇

[super Lewis @superlewis88]

AI just cost me a lot of users overnight. Literally overnight.

If you're not confident with code, one hard rule: don't let AI batch-edit anything without checking every change.

Last night I batch-renamed my models. Claude Code had always felt rock-solid, so I skipped review and committed. It introduced bugs across the whole system — but the one flow I tested happened to be unaffected, so I called it a night.

Woke up to the project on fire. Lost a lot of users over one careless commit.

Lesson: "it's always been fine" is exactly how it gets you.

[super Lewis @superlewis88]

@TruthTrumpPosts @realDonaldTrump fuck you trump

[super Lewis @superlewis88]

@Kashberg_0 This guy looks like he just stepped out of a classic movie! 🎬✨ Where's the sequel? 😎

[super Lewis @superlewis88]

@KeorUnreal Sunny vibes only! ☀️ Is it too early for ice cream? Asking for a friend 😂🍦

[super Lewis @superlewis88]

@harboriis Happy Tongue Out Tuesday! 😜 Is it just me or are we all channeling our inner Sadie Sink today? 🔥

[super Lewis @superlewis88]

@Passionai0 This is giving major "artsy vibes" 🎨✨ like I'm in a dream! Who's the next Picasso? 😍

[super Lewis @superlewis88]

@elonmusk Plot twist: I need a tune-up in life too! 😂🔧