A single-founder build means every one of these was found, diagnosed, and fixed solo. Shown here in full because the reasoning is the actual work, not the parts that happened to go right on the first try.
Bug Report
The recommendation engine ignored what people actually typed in
A tester wanted the same amount every time toward a $1,000 to $1,500 goal. The #1 recommendation was a $300 challenge with a varying amount every save, the opposite of both numbers entered. The matching logic behind the scenes added up several different signals to pick a winner, but nothing made sure the ones that actually mattered outweighed the ones that didn't. Getting the dollar amount close barely moved the score, saying "same amount every time" and getting handed a challenge that changes every time cost nothing at all, and checking off a few unrelated complaints could quietly add up to more influence than either of those.
How much the dollar amount mattered: barely → the single biggest factor
Picking a style that didn't match: no penalty → now counts against it
Stacked complaints boosting a bad match: unlimited → capped
Deeper Bug, Found By Auditing The Fix
The first fix wasn't actually the fix
Just making the dollar amount matter more looked like the fix, until every single combination the quiz can actually produce got tested by hand, around 64,000 of them across all four saving styles. Three of those styles, gradual, random, and mixed, were still getting matched to a challenge that couldn't actually deliver what they promised, 38 to 62 percent of the time depending on the style. One rule fixed all three at once: if someone picks a style, a challenge that doesn't genuinely match that style gets knocked far enough down the list that it stops winning by accident.
Every answer combination checked by hand: 64,000
How often it still got it wrong after the real fix: almost never
Coverage Gap
39 challenges, not 10, because the engine needed real options
The scoring fix could stop a bad recommendation, but with only 10 preset totals, plenty of legitimate amount and frequency combinations had no genuinely good match to point to in the first place. "I want a challenge where I can save a thousand dollars with just $1 bills" became the literal test case: 29 new presets across fixed-amount, gradual, random, and variety styles, spanning $1 to $500 per save, so every quiz combination has a real answer instead of a compromise.
Ready-made savings plans to choose from: 10 → 39
Results Page Overhaul
Standardizing time revealed a tradeoff nobody could see before
Some challenges described their pace in weeks, others in raw save counts, with no consistent way to compare them. Giving every preset a real cadence-days field and computing an honest duration surfaced something previously invisible: the exact same $1,000 goal takes 2.7 years with $1 bills but 12 months with $20 bills. The results screen also stopped showing every one of the 39 presets regardless of fit, gating anything below 70% of the top score behind a "look for more options" tile instead of presenting a weak match as a confident answer.
$1 bills to $1,000: 2.7 years
$20 bills, same goal: 12 months
Design Decision
The icon redesign measured what "bigger" actually meant
The original icon was a flat green square, technically fine and forgettable. The redesign asked for a light-green gradient with an enamel-pin feel, checked against real contrast math across three intensities rather than eyeballed, since a pale gradient and a white glyph fight each other once the top of the gradient gets too light. The bigger finding came after color was settled: the glyph itself only filled about 70 percent of its own canvas. Once measured properly, the fix was obvious, scale it up to roughly 85 percent visible fill, close to what most icon-heavy apps already do.
Shipped, Then Reverted
A hard drop-shadow on every icon, live for about a day
Trying to give the flat app more personality, every icon got a hard offset shadow behind its glyph, a one-line CSS rule that touched every icon in the app at once. Seeing it live for real was enough: it looked bad, and the call was immediate, revert it. A simple stroke-width increase on the icons stuck instead, bolder without the shadow gimmick, and without touching the dynamic per-theme recoloring the shadow version would have complicated.
The Same Lesson, Learned Twice
Flat cards got real depth, and dark mode broke the same way twice
The whole app had been intentionally flat, a single pixel border and solid fill on every card and button. Soft gradients and shadows fixed that, built from the app's own color tokens so the effect holds up across all six theme combinations. Both times this kind of highlight got added, mixing it toward literal white looked fine in light mode and quietly blew out contrast in dark mode, where a background that was never meant to get that light suddenly did. The fix, both times: mix toward a token already calibrated for the current theme, never toward a hardcoded white or black.
Geometry Bug, Caught Before Shipping
A phone-mockup tool that worked at a steep angle broke at a mild one
Building tooling to composite real screenshots into device mockups, a technique using a mask's four extreme pixels as the four corners of the screen worked perfectly on a dramatically tilted iPhone frame and produced sheared, unusable garbage on a mildly tilted one. The extreme-pixel trick only holds when a rotation is steep enough that each corner is uniquely extremal, for a near-upright rectangle the leftmost pixel sits mid-edge, not at a corner at all. Switched to a proper minimum-area-rectangle calculation that finds the true corners at any angle, verified against both the broken case and the original working case before trusting it.
Design Judgment, Not Just a Fix
A geometrically correct mockup still wasn't a good one
After the corner-detection bug above got properly fixed, the resulting tilted mockups were technically accurate and still didn't look good enough to ship. Reverted to the flat, plain treatment that was already working, and later replaced with mockups built by hand in Figma instead, which were immediately, noticeably better. Correct geometry and a real product photo turned out to be two different bars, and only one of them was worth shipping without more design work behind it.
