Recordário

Situation

A product that doesn't exist until the customer makes it

Recordário turns customers' photographs into watercolour fine-art prints: framed art, pet portraits, mother-and-child compositions. I designed and built all of it alone.

Two constraints shaped every decision.

The product doesn't exist until the customer makes it. There is no catalogue photograph that tells the truth. Every unit is generated from the customer's own photos by an image model, takes roughly ninety seconds, and can fail. The patterns e-commerce hands you, meaning product photography, reviews of this item, and a straightforward add-to-cart, don't map onto a product that is authored at the moment of purchase.

One person, no team. A functioning storefront needs content, paid media, creative, analytics and operations. All of it competes directly with building the product.

And then the real device budget: 88% of traffic is Android, much of it mid-range, much of it arriving inside Instagram's in-app browser. Not the machine the design was drawn on.

Task

Two design problems, deliberately treated as one system

Make the customer the author. Move the moment of ownership before the transaction. Design a flow where a nondeterministic, ninety-second, occasionally-failing generation reads as craft rather than as a loading screen.

Build the operating system for a company of one. Encode brand voice, product truth and persuasion strategy as machine-readable context, so agents could produce content, creative and analysis at volume without diluting taste.

Before building the modal that catches people leaving with unsaved art, I shipped only the measurement: a passive event that records the leaving, changes nothing about it, and runs only while the retention feature is switched off. I wanted the true size of the leak before I built the plug, and a clean baseline to judge the plug against afterwards.

That habit, instrument first and build second, is the one I would most want to be judged on, so it belongs at the top rather than buried in a later section. It is also the habit that caught the two problems in the reflection at the end, both of which are mine.

Action

1. The preview is the product, not a step toward it

The endowment effect is the entire strategy. People overvalue what they have made. So the flow hands the customer authorship before checkout: a live canvas on the real fine-art paper texture, their photograph draggable and pinch-scalable, their message repositionable, physical frames and in-canvas frames swappable live. The generated watercolour sits under a multiply blend on the actual paper, so what they see is what ships.

The consequential decision was to make generation free, generous (five per day), and account-free. Each image has real cost. Gating it behind an email would have protected that cost and destroyed the mechanism, because the preview is the conversion event, not a lead magnet.

Instrumented outcome: customers swap the in-canvas frame about thirteen times per preview. I read that as engagement for months. I no longer do, and I come back to it at the end.

2. Designing the ninety seconds before the reveal

The reveal is the peak the whole visit is built around: the watercolour arriving on the real paper. It is also the most fragile moment, because the generation takes about ninety seconds and can fail. That turns the wait itself into a design surface. How you spend a customer's patience decides whether they ever reach the peak.

A generation that succeeds is nearly always done by around 110 seconds, so waiting longer buys almost nothing. The real decision was where to spend the patience I had.

So rather than making the customer watch a stalled bar, the product cuts its losses early and quietly starts the painting over. Pet portraits are the core product, so when one provider began refusing about 12% of pet photographs on content grounds, those go straight to the model that will render them.

The reveal that was failing for roughly a third of customers now lands. Zero generation errors across the thirty-two daily reports since the fix, and on the most recent measured day, 31 generations produced 31 previews.

3. Designing for the device the customer actually holds

Customers were losing their uploads from the third photograph onward. A composition takes four. So the product was breaking at precisely the moment someone had committed to it, and breaking hardest on the mid-range phones most of the audience actually holds.

The fix itself was mundane, so the number is what matters: on the most recent measured day, 44 of 44 upload sessions completed with zero friction. The customer reaches the canvas with all four photographs intact, on the phones where the flow used to break.

The drag gesture took three documented rounds against real device evidence.

Detail work on top: a magnetic centre snap that takes 3% of drift to catch you and 5.5% to let you go, and the asymmetry is the entire reason it feels magnetic rather than sticky. Haptics on snap and on drop. And the gesture runs outside the app's normal rendering path, because updating state on every frame made the drag stutter on exactly the mid-range Android phones most customers are using.

4. Persuasion as a design system, not a copywriting mood

I wrote a persuasion constitution that governs all four channels: ads, social, blog and in-product writing. It's grounded in Ariely and Kahneman, and it starts from one uncomfortable premise about who is actually doing the buying.

System 1 buys. System 2 only looks for reasons not to. In an emotional gift purchase the fast, automatic, feeling mind makes the decision, and the slow rational one arrives afterwards hunting for objections: price, delivery time, risk. Two rules fall out of that and run through every screen. Emotion comes before information, so a price never appears before a feeling. And what Kahneman calls WYSIATI: if the emotional benefit isn't visible on the screen, it doesn't exist for the decision. An unstated benefit isn't a subtle one, it's an absent one.

From there the document is a working arsenal rather than a manifesto. Twelve biases, each with a named job in the product.

Then each block of the funnel gets one target feeling and one master bias, so any piece of content, mine or an agent's, knows what it is for before anyone argues about wording.

What makes this a design system rather than a mood is that every one of those decisions has an event behind it. The frame swap, the room scene the customer zoomed into, the exact stage they abandoned at, the moment they walked away with finished art still unsaved. The constitution says which bias to pursue. The event catalogue says whether it actually worked.

Its governing rule is the part I'm proudest of, because it's a separation of concerns.

The brand landmines are specific and absolute: no em-dashes in Brazilian copy, never claim "handmade", emoji parsimony, a call to action matched to the funnel stage. In a conflict, the brand wins and the bias loses. That constraint is precisely what made it safe to hand copy to agents without the brand rotting.

The constitution carries a second list, and it turned out to matter more: the anti-biases, the moves that actively destroy conversion. Price before emotion. Latency with no feedback. Delivery time framed as logistics. Invented scarcity, which System 2 detects and never forgives. A lukewarm ending. A gate that blocks the button on someone who has already invested. And choice overload.

That is choice overload, in exactly the place my own document warns about, inside the product that document governs. The system was right and the screen was wrong. Writing the rule down did not protect me from breaking it. Instrumenting it is what caught me.

Which leads to the only honest way to close this section: stage by stage, what is the evidence actually saying?

Where the funnel actually breaks

The pattern across that table is the thing I'd want to be judged on. The parts of the system I could specify, I got right. The part that depends on reading evidence honestly is the part that is still open, and it stayed open for six weeks while my own daily report was flagging it.

5. The operation a company of one can actually run

One person cannot write the content, run the paid media, produce the creative and read the analytics and also build the product. Each of those hours is taken from the others. So the operation itself is a system: fifteen agents reading a shared hierarchy of context that encodes brand voice, product truth, the persuasion constitution above and the full event catalogue. Two of those documents regenerate from the production database on every build, so the agents' picture of the product never drifts from what customers actually did.

The one that carries this case is the daily analytics agent. Every morning it pulls the site funnel and the previous day's ad spend, works out where conversion moved, writes a dated report, and proposes the next test to run. Thirty-eight consecutive reports. It is the loop that surfaced every problem in the reflection below, including the 84% conversion cliff and the blind spot I built myself.

The same fifteen agents also carry the content and media operation, none of it competing with product time:

• Editorial pipeline. Five agents chained from strategy to publish, straight to the production database. 37 posts live.
• Paid media. Meta and Google Ads managed through agent tooling.
• Social creative team. Visual-first, so the art director leads and a brand guardian holds the brand line.

One more decision worth showing is the autonomy gradient. Each agent's leash is set by the blast radius of its actions, not by a uniform policy.

6. An evidence loop, not a dashboard

Every event in the product is catalogued, and the agents interrogate it live rather than read a frozen chart. They read production, but only read: they never see a customer's personal data, and the keys that could do real damage never sit on my laptop. That is the difference from a dashboard. A dashboard reports what happened. This loop proposes what to do about it.

Result

Ratios and relative changes

Absolute revenue withheld. Read these as the floor the design now stands on: the customer reaches the reveal every time, which makes the conversion gap in the next section a design problem, not a reliability one.

Operating leverage, one person

Paid media efficiency

Eighteen-day window, platform-attributed. The site doesn't yet carry campaign parameters, so the campaign-to-conversion crossover is directional, and I label it that way in my own reports.

• Cost per lead ~R$2.19, best campaign R$1.29 at 6.2% CTR
• Best single day 8.8× ROAS
• Blended across the window 0.88×. The channel is not yet profitable at blend, and I know it.

Live and transacting on both instant bank transfer, with a transparent on-site QR flow, and card.

Reflection

The three things I'd put in front of a design team first are the ones that aren't solved

The 84% cliff

Thirty-one previews produced five purchase-intent clicks. The preview works. The transition from preview to object does not.

My hypothesis is that the endowment effect I engineered creates ownership of the image, not of the thing. The customer already got what they came for. Next test: make the physical object the hero at the moment of approval, at scale, on paper, on a wall, instead of presenting a size dropdown.

What the customer is actually buying

Thirteen frame swaps per preview

I read this as engagement for months. I now think it's Hick's law, choice paralysis arriving at exactly the wrong moment in the funnel. Next test: pre-select a recommended frame and demote the rest to a secondary affordance.

A blind spot I built myself

One funnel event has reported zero for six weeks while the steps on either side of it show traffic. It's broken instrumentation sitting directly on the drop I most need to explain, and my own daily report flagged it independently at both ends of that window while I didn't act on it. The lesson isn't "instrument more".

One more, on judgment rather than metrics: the social pipeline is architecturally complete and barely exercised, one finished post against twenty-four editorial pipeline runs. I built it ahead of need. Given the same seven months again, that capacity goes into the 84% cliff instead.

What this case is evidence of