One Brief, Eleven Platforms

The first time I ran Finfold end to end, the screen filled neatly with eleven pieces: a LinkedIn post, an X thread, a Xiaohongshu card set, a video script, a newsletter. As a demo, it already looked sellable.

I didn't celebrate. Because I knew something about those eleven rows that nobody watching the demo could: a few of them were never generated by the model at all. An early build had a "smart" fallback — the moment a model call errored out, it quietly slipped a prewritten template into the empty slot and moved on. What the user saw was a clean row of eleven. What I knew was that some of those slots were stage dressing.

That bugged me for days. Then I worked out why: the real trap wasn't a failed generation. It was a mediocre template dressed up as "AI success." Once a user trusts it, takes it, and ships it, the thing they break isn't a quota — it's the account they've spent years building.

The first feature I killed with my own hands: "always succeeds"

Silent fallback is a seductive trap. The demo never breaks, the audience nods, you feel good. But it has a fatal flaw: the user can't tell which slot is model output, which is a cached result, and which is a template smuggled in by the backend. Send out one piece with a wrong fact in it, and you've wagered a three-year-old account's credibility against a single free retry. That math doesn't balance.

I sat with it for two days. Then I deleted it.

Here's a piece of product common sense a lot of people miss: an error message isn't an engineering band-aid — it's part of the user's decision. A dry "generation failed" says nothing. Swap it for "the Xiaohongshu image caption ran over the limit; the other ten are saved — want to retry just this one?" and the user finally knows the next move.

Why I cut the all-purpose chat box with my own hands

The earliest version had a chat box too. Conversation on the left, results on the right. Type "turn this brief into eleven platforms," wait for the answer. The structure felt effortless — because every AI product does it this way, users know it, and copying it is the path of least resistance.

The moment I pressed a real operating scenario against it, it cracked.

People doing the actual work don't work like that. They lock one confirmed figure and only rewrite the headline. They compare two X openings side by side while sneaking a glance at whether Xiaohongshu has any banned words. They shelve one platform until next week. A chat log can store your conversation, but it can never express, for eleven objects, what version each one is, what state it's in, and which depends on which. Managing that in a chat box is like running a project in Notepad — it works until it doesn't.

Guardrails belong at the door, before generation runs

I stepped in this hole too: brand checks used to sit at the very end. Eleven pieces fully generated, then the system would slowly surface "tone might be inconsistent." Was it right? Right. Useful? Useless. The user had been editing for half a day — and now you tell them the whole block can't ship?

So I moved the whole thing ahead of generation. The guardrail holds disallowed promises, mandatory product names, figures that can't change by a single character, tone boundaries, and CTAs. When generation runs, these rules enter the context first; after generation, they get checked again. This is closer to an executable institutional memory than a brand manual gathering dust in some cloud drive.

Four seconds of waiting shouldn't be just a spinner

Eleven platforms generating at once can't be instant. The early version just spun a circle on the button — four seconds that, in the user's eye, stretched into fifteen. Later I had each card light up its own state: reading sources, applying rules, generating, checking — and each finished card surfaced the moment it was done. Total time barely moved. But the wait went from a black box to a process you could actually watch.

I did not write that up as "perceived speed up 40%" — because I never ran the experiment that could support a percentage like that. What I can vouch for is two things: users can see where the system is stuck, and they can go handle the ones that are already done. The first keeps them calm; the second keeps them from waiting dry.

The pricing page forced me to answer the question I was most afraid of

Building the pricing page, I got stuck. If a product charges only by "how many generations per month," it sooner or later gets dragged into a token-resale price war — whoever's model is cheapest wins, and the product you painstakingly built has a moat of roughly zero.

What's genuinely hard to copy is a different pile of things: reusable brand rules, accumulated historical context, cross-channel state management, a publishing rhythm. So pricing can't just sell quantity — it has to line up with the way of working you actually provide.

Honestly, I still don't have enough real paying-user data to prove these tiers are the right ones. That's the thing I feel least sure about right now. I don't think the next move is to pile on features — I think it's to go back and stare at one question: when a creator pays, what are they actually paying for? Saving time? Avoiding mistakes? Or keeping the brand in one consistent voice?

In the end, this product turned around and taught me how to make content

The interesting part is this: after building the product, I also figured out, for myself, how I publish. The source lives on my personal site — the full argument, the images, the sources, all here. LinkedIn distills the professional judgment. X logs the short, hard-won conclusions from the work. Xiaohongshu takes one mechanism and breaks it into a card set. The video channel puts one concrete scenario into plain language. They share one set of facts; none of them copy each other's sentences.

Take this very article. On Xiaohongshu I won't compress it into "5 principles for AI content tools." I'll talk about one thing: why I deleted a feature that could have kept the demo succeeding forever. On X, it's a one-line verdict with four screenshots of failed states. On LinkedIn, the angle is why AI products have to put their uncertainty on the table. The site is where the argument gets made in full; the platforms are where the people get walked in.