Shirley XueAll work
ablefyPrototypeAI-assisted2026

A pricing form that advises, then you edit

Problem
Sellers couldn't find the price, and the form asked a billing-mechanics question no seller could answer. Optional settings came first; the checkout preview sat at the bottom of the page.
Solution
First a reorder with an always-on preview. Then, when consumer-credit rules retired one plan type, two questions a seller can answer instead of a plan-type picker — and an AI advisor whose every number comes from a deterministic engine.
Result
Five front-door prototypes, a comprehension test ready to run with its decision rule already written, and a three-phase plan in front of stakeholders and engineering. In September 2026 the first piece, a buyer-commitment control, was designed for the production form. Nothing is in production yet.

Role & team

Design and front-end, built solo

  • Designed every layer and coded the front end, directing AI coding agents.
  • Validated with stakeholders as working prototypes; the August redirect is a written proposal.

Duration & scope

Audited 2023 · rebuilt 2026 · redirected Aug 2026 · first piece for the live form Sep 2026

  • The 2023 audit covered the plan booking and upgrade flows; a 2025 plan-booking redesign carried the first fixes.

Receipts

No invented numbers, and a gate that keeps it that way

  • Checked across an 85-case automated test and two live runs.
  • All 44 production settings accounted for. One legal cap found unenforced; the prototype now blocks the save.
Context

At ablefy, a creator-commerce platform, every sale starts in one form: pricing model, price, payment methods, schedule, trial. I rebuilt it in layers. In August 2026, consumer-credit rules retired one plan type, and the design question moved from how to lay the settings out to what to ask the seller.

Before

The form sellers struggled with

A design audit and two feedback rounds found one pattern: sellers couldn't find what matters, got too much too soon, and didn't understand the words.

The old pricing form — toggles and optional settings above the price field, payment methods in two separate fields, checkout preview at the bottom of the pageView full size
The old form (staging) · options before price · preview at the bottom
  • Pricing model, payment, currency, schedule and trial all had equal visual weight: no hierarchy.
  • Optional toggles appeared before the one mandatory field: the price.
  • Payment methods lived in two redundant fields (“preferred” vs “available”) with no explanation.
  • Plan-type names confused people: subscription vs. limited subscription vs. installment.
  • The checkout preview sat at the bottom, so checking your work meant scrolling on every edit.
Layer 1

Early 2026 — reorder the form around the order a seller decides in

The first fix was information architecture before visual design. The checkout preview also moves beside the form, always visible and updating as you type.

The same controller engine powered a low-cost spin-off: an offer builder for sales calls. Its editable per-payment timeline rebalances the other payments when one changes, so the total never drifts.

The redesigned pricing form prototype — pricing model cards first, then amount and schedule, with a live checkout preview panel fixed on the rightView full size
The redesigned form (prototype) · flow reordered · checkout preview always in view
  1. 01

    Pricing model

    The four plan types the form had then, moved to the front, because everything below depends on this choice.

  2. 02

    Price & schedule

    The mandatory part, no longer buried under toggles.

  3. 03

    Payment methods

    One consolidated field with logos, ordering and a preferred method.

  4. 04

    Optional settings

    Trials, custom names, availability: after the essentials.

Review

The editor I proposed, and what the review changed

The second layer let power users edit the plan directly on the checkout preview: change the thing you're looking at, not a form that describes it. The idea survived; my first shape for it did not.

What I proposed

A standalone editor page: the checkout as the whole surface, every plan setting editable in place, the form left behind.

What the review said

A second place to edit pricing would split the seller's sense of where prices live, and any disagreement between the two surfaces would have to be fixed by hand. Rejected, rightly.

What I changed

An overlay on the same redesigned form, sharing one state underneath: one place prices live, two ways to look at it. The review accepted it.

What it cost

Rebuilding the prototype's controls on the platform's own components, instead of one-off colours that would drift from production.

Reframe

August 2026 — the form was asking the wrong question

Three things arrived at once. Consumer-credit rules, applying from 20 November 2026, retired the installment plan as a seller choice. Engineering asked for a refactor of the pricing services instead of another re-skin. That removed the constraint behind every earlier prototype: a front-end-only change that saved exactly what the old form saved. And the requirements document named the real problem: the form asks a billing-mechanics question when the seller only has a business answer. Its fix was three better names on the same picker.

I argued that new names cannot fix a recognition problem, and proposed changing the question instead: two questions a seller can answer, in this order. The plan type then becomes a result the seller can read and override, not a choice made without context. Underneath, one saved offer profile that a manual, an assisted and an AI version fill to different depths.

Roughly nine in ten installment orders used one configuration, so the form's other settings were noise for almost every seller. And in 205 recorded seller calls the word “installment” never occurs; the people explaining it say “payment plan”.

An earlier rename, “limited-term plan”, did not survive; the product calls the type a Program. Its legal basis did: whether content is delivered once or over a period decides both the VAT treatment and the credit rules, and that became question one.

The front-door prototype inside a recreated cabinet shell — step one asks when the buyer gets the content, with “Piece by piece, while it runs” selected; below it the money question offers the full amount up front or as you deliver; on the right a buyer-view panel fills in as you answerView full size
The two questions (prototype, synthetic test product) · the buyer's view fills in as you answer

1 · When does your buyer get the content?

All of it at purchase, or piece by piece while it runs. This decides which law applies, so the seller answers it and the system never guesses.

2 · When do you want the money?

The full amount up front, or as you deliver. A business decision with both answers allowed, and the only place assistance belongs.

Doors

Five doors, one test, and a rule written before the sessions

I built five front doors on one shared offer profile, each testing a different seller strategy: choose the shape, describe the offer in two questions, pick the closest template, start from the money, or say it in one sentence. Two earlier variants were cut as variations of one door (one hid a label, one split the same two questions across pages); testing them would have measured copy, not a design idea.

Before any seller saw the copy, a synthetic seller panel (six personas grounded in real seller voice history, treated as hypotheses only) read the first question as how long the course lasts, not when it unlocks, on the most common product shape. That answer would have been legally wrong. The question now asks about unlocking, and the door that guessed a delivery shape from the word “course” no longer guesses.

The comprehension test compares the form with the buyer's checkout as the editing surface: between subjects, three sellers per arm. The decision rule was written before the sessions, so the team's preference cannot decide the result; a tie ships the smaller change. Round one has not run yet.

Door B after both questions are answered — the answered step folds to one row, a green card names the result “Program — 3 × €149.00 monthly, €447.00 total” with a fixed-length or runs-until-cancel switch and a one-line reason, and the checkout preview on the right shows what is due todayView full size
Door B, answered · the plan type arrives as a named, explained, switchable result, with the checkout beside it
The other test arm — the buyer's checkout page as the editing surface, with a side rail listing offer, price, schedule, run length and payment methods, and the checkout itself showing what is due today and the payment methodsView full size
The other test arm · the buyer's checkout as the editing surface, with the same two questions
Door D, money first — after the seller enters 3 monthly payments of €150 and says everything unlocks at purchase, a card titled “You cannot split this price” explains that splitting the price of something already delivered is lending, converts the schedule to one payment of €450, names Klarna as the buyer-side route, and states that plans already sold are not affectedView full size
The refusal (door D, money first) · the seller's numbers carry into one payment, the buyer-side route is named, and the card says existing plans are unaffected
Settings

No setting dropped by accident

Simplifying a form risks dropping, by omission, the one setting a power user's business runs on. So every control the live form renders was read from the production code with its real limits and classified, each with its evidence and an owner. The rule: simplify by defaults, never by removal.

The inventory also found a legal gap. Consumer law caps a standard-terms commitment at 24 months; the field was uncapped, and live plans ran past it. Legal confirmed the cap as a hard limit. The prototype now blocks the save whenever the number of payments times the interval exceeds it, because 36 monthly payments bind a buyer exactly like a 36-month minimum term.

Door B with the number of payments set to 36 — the result reads “36 × €12.42 monthly”, the footer says “36 months binds your buyer past the 24-month cap — reduce the payments or stretch the interval”, and the Save plan button is disabledView full size
36 monthly payments · the commitment passes the 24-month cap, so Save is disabled and the footer says why

44 production settings44 in total, every one accounted for

  • 26 built · carried into the design, first-class or behind a working default

  • 6 moved · move out of the form to a named place; promotion settings go with promotions

  • 10 open · undecided, each with a named owner; the prototype may not resolve them by accident

  • 2 dropped · removed on purpose, with the reason on record

Counts from the requirements inventory, September 2026. An automated check in the prototype re-counts them on every build, so no setting can drop out unnoticed.
Advisor

An advisor, not an authority

In the third layer the seller states a goal instead of configuring a form, and the advisor proposes a pricing structure. Since August it runs on real order data, pseudonymised, and opens from the product the seller is looking at, not from a menu.

The question was never whether the advisor could produce a plausible answer, but whether a seller would know when to trust it, since these recommendations move real money. One failure would be fatal: a made-up number. So one decision was fixed before the advisor existed: separate the recommendation from the outcome.

The language model proposes and scores options, argues trade-offs and writes the rationale; a deterministic engine computes every number, filters what is feasible, and arbitrates what the model may claim. If the two disagree, the engine wins. No recommendation was shown to anyone before it was tested for exactly that failure: across an 85-case automated test and two live runs, no invented number appeared. On every change, 12 deterministic graders and type checks must still pass, so it is a standing gate, not a one-time score.

The pricing advisor prototype inside a recreated cabinet shell — a recommendation card for a pseudonymised seller's membership, “Add a 6-month minimum term”, with the plan shown as the buyer would meet it; below it a basis block where each row names its source: Your data, Computed — “Six months is a judgement, not a calculation” — and Platform avg; then a “What pressing this changes” panel listing now, after, stays and always; on the right, the buyer view of the same proposalView full size
The advisor (prototype, real order data, pseudonymised) · a recommendation, the basis of each figure, and what pressing it changes
The advisor drafting a price for a new PDF guide — one payment, with a basis row reading “You said the buyer gets everything at once — so the price cannot be split over time here. That is the product deciding the shape, not the advisor.”, a price range drawn from the seller's own catalogue, and a note that a name alone is only evidence the seller confirmsView full size
The refusal · the guide is delivered at once, so the advisor will not split its price
  1. 01

    State a goal

    Cash flow, conversions or completion, in the seller's terms rather than the form's fields.

  2. 02

    Read the recommendation

    A proposed pricing structure with a plain-language reason, or “nothing to change”, an allowed answer, so the advisor is not pushed to change prices for the sake of it.

  3. 03

    See the consequences in euros

    What the seller receives now and what the buyer sees, beside the recommendation, before any decision.

  4. 04

    Read where each figure came from

    Each figure shows its basis: the median of the seller's own orders, a platform benchmark, or a judgement labelled as one.

  5. 05

    Change anything

    The advisor drives the same plan model as the manual form, so every field it filled stays editable by hand.

  6. 06

    Save the plan

    The seller saves it after a summary of what changes now, what changes later and what stays untouched, with an undo.

Reversals

What I reversed in the AI layer

The advisor changed more between June and August than any other part, each time after it was wrong in a specific way.

The June glass-box view of the pricing advisor — a node graph from seller goal through feasibility filter, consequence bundle, judge scoring and critique to the final recommendation, with a panel showing one node's inputs and outputsView full size
June 2026 · the retired glass-box view: every node clickable, a live-looking window over a recorded run

The glass box, retired

In June every node of the decision was clickable, a pipeline view that read as proof. But it showed recorded runs, and a live-looking pipeline implies a live computation, so it promised too much. It left the product. Each figure's basis now sits on the figure itself, where the seller is already looking.

The product decides the shape

The advisor proposed splitting the price of a download, because its guard read the seller's history, not the product. Splitting the price of something already delivered is lending, not a payment plan. The fix: one source for the delivery fact, read by both the wording and the rules.

Compliance is the minimum, not a goal

Compliance sat in the goal row beside cash flow and conversions, which implied a trade-off that does not exist. It left the row; the rules apply whatever the goal.

What the advisor may learn

A valid use case becomes advisor knowledge; a workaround becomes a design fix and is never taught to the advisor.

Sep 2026

September 2026 — the first piece, designed for the production form

While the redirect waits on the organisation, one part could go ahead. The plan type that survives the credit rules needs a buyer commitment: a minimum term the buyer cannot cancel early. The production form already had a field for it, and since October 2026 the backend enforces the 24-month cap on limited subscriptions.

Today the seller types a number and a unit (“6 · Months”) with no link to the payments the plan charges, and the checkout tells the buyer nothing about the commitment. The design states it in payments: a “Buyer commitment” switch, last in the form's switch list, and beside the price a select reading “Buyer commits to all 6 payments” or “the first 3 payments”. The duration is derived from those payments, saved in the field the backend already has, and cut at 24 months. The line under the select shows what the buyer will read: “6 months, no early cancellation”. It replaced an earlier seller-side draft, “cancelling takes effect after the last payment”, a euphemism for the same thing.

A chosen “first 3 payments” stays when the seller changes the count; an untouched “all payments” follows the plan. An existing plan saved with a term past the cap gets no error. The select shows what is saved and what saving will do (“Currently 48 months, longer than the plan; becomes 24 months when you save”), because the seller did nothing wrong; the field did.

The production pricing-plan form for a limited subscription, recreated: a Minimum term switch on the right, and under the price fields a number box reading 6 and a unit select reading Months — a term typed with no link to the six monthly payments above itView full size
Today (production, recreated) · a number and a unit, not linked to the payments
The same form with the proposed control: a Buyer commitment switch last in the switch list, and under the price fields a select reading “Buyer commits to all 6 payments” with the line “6 months, no early cancellation.” beneath itView full size
Proposed (prototype on production's components) · the commitment in payments, with the buyer's line beneath
The proposed control on an existing plan saved with a 48-month term over 24 payments: the select reads “all 24 payments” and the line beneath says “24 months, no early cancellation. Currently 48 months, longer than the plan; becomes 24 months when you save.”View full size
An existing plan past the cap · a notice instead of an error
Status

Where it stands

The redirect is a proposed three-phase plan, in front of stakeholders and engineering: the front door and the saved offer profile first, then the same screens pre-filled from what the product already knows, then the advisor. Migration is decided: running orders finish, plans that no longer conform close to new buyers and stay alive, new plans use the new design. The first piece to reach a production ticket is the buyer-commitment control, filed beside the backend change for the cap.

The comprehension test is ready to run; round one waits on recruiting. The front-door and advisor prototypes are standalone builds. Nothing here is in production yet.

What carries beyond the prototypes: the two questions, the settings inventory and the legal cap it found, the rule that every figure shows its basis, and one lesson from the AI layer: trust in AI judgment comes from the deterministic engine arbitrating, not from a better prompt.

In parallel I led the platform's charging-model research: a cross-team demand inventory, a build-vs-buy scan of the billing market, and a usage-based pricing recommendation. It is the strategy layer this form serves.