One summer. 104 matches. 20,451 predictions.

The 2026 World Cup was the biggest ever staged: 48 teams, 104 matches, hosted across the United States, Canada, and Mexico. The prediction games built for it split the group stage from the knockouts, buried the fun under ads, or treated playing with your actual friends as an afterthought. My bet: one clean bracket, built around private friend leagues, could turn casual viewers into a summer-long ritual.

This was a project built for fun, passion, and learning, not money. Free to play, no ads, no paid tier; the only revenue surface is an optional tip jar. That freedom set the quality bar: every decision could optimize for the player.

First commit May 1. First kickoff June 11. The six weeks in between were nights and weekends around a full-time role: feature-complete for a friends beta about a month before kickoff, then hardened for strangers in the final stretch. What shipped: 47 routes in 12 languages, an installable PWA, one-tap Google sign-in, the 104-match bracket with its cascading knockout tree, bonus picks, private leagues with invite links, live leaderboards, a share-card studio, an AI news desk, a public stats hub, and a penalty-shootout mini game standing in for the 404 page.

A design system, not a pile of screens

Every surface draws from one hand-rolled token catalog: an OKLCH palette, type scale, and status colors in a single theme block, consumed by a small shared primitive layer. Cross-surface consistency is a written contract, not a habit: one documented pick-vs-result grammar covers every place a pick is rendered, and a unit test pins the shared treatment so surfaces cannot silently drift.

Shared data, private by construction

A social game runs on shared data, so privacy is structural, not polite. Row-Level Security is forced on all 21 tables; social surfaces share counts, never values (a league roster shows "98 of 107 picks", never the picks); and every table’s exposure is declared in a manifest that CI verifies against the real schema on every change. A future competition cannot inherit the World Cup’s visibility rule by accident: the rule is a required field with no default. 148 adversarial probes and a weekly fresh-context audit kept the whole thing honest.

Pressing share

The hardest ship of the six weeks was not technical. Posting a personal project to LinkedIn, something made for the love of it rather than for work, took real vulnerability. I hit publish in late May, once a two-day adversarial security audit signed off the product as safe to promote to strangers.

Growth was engineered around one moment: the bracket lock at first kickoff, June 11 at 19:00 UTC. A five-email countdown drip anchored to that deadline, invite links that unfurled into a league preview, and a printable schedule people could stick on the fridge all pointed at the same date. 60% of all players joined in kickoff week.

The deadline did the selling. Over the final day before the lock, completed brackets surged from 94 to 146, up 55%, with the last-call email converting 32% of the players it caught mid-bracket into finishers. The last bracket was completed eight minutes before kickoff. Across the whole campaign the system delivered 1,034 emails with zero spam complaints.

The 15-minute budget

The product’s depth is also its steepest ask. Footy26 wants an exact scoreline for all 104 matches: that is what makes every match matter and every leaderboard swing exciting. But in the expanded 48-team format, it is a real demand on a passerby’s attention. I treated about fifteen minutes as the ceiling a curious visitor would give the game, and spent design effort against it: two taps per pick with an empty save defaulting to 1-0 (visibly hinted), auto-save in the visual editor, and resume-anywhere progress.

From kickoff to the Final, Footy26 ran on the tournament clock: scores ingested every minute, leaderboards moving during matches, the knockout tree resolving round by round against 146 frozen brackets. Sessions averaged nine and a half minutes at a 22.5% bounce rate, and the spectator surfaces out-drew the editor: more players watched league standings move than ever touched their own bracket again after the lock.

Stoppage time on day two

Operating started immediately. Two days in, some dashboards were hanging on their loading skeletons: a rare race in the framework’s streaming hydration that only live-traffic timing could trigger. The response became the playbook. Bounded self-healing shipped before the root cause was known, turning a stuck page into a slow one; live probes then read React’s internal fiber state inside stalled production tabs until the race had a name; and every closed defect left a guard test in CI so the class cannot recur. Contained in 48 hours, cured for good by a framework upgrade weeks later, and the diagnostic playbooks are now part of the repo.

Shipped on signal

The live window is where instrumentation earned its keep. Every user-facing flow emits an event, so when behavior surprised me, the product changed while the tournament was still running:

The invisible last step

01

1. Signal

   Players finished all 104 match picks, watched three of four bracket cards flip to "Complete", and stopped, never discovering the three bonus picks.

2. ↓→

   Shipped

   A prompt at the exact moment the 104th pick saves, a bonus status card on the summary, and a persistent "All 3 bonus picks saved ✓" state replacing a dead grey button.

3. ↓→

   Result

   The prompt converted 87% of the players it reached. Bonus completion (85% of builders) ended ABOVE full-bracket completion.

As shipped

Picks at a glance

02

1. Signal

   The #1 in-tournament question was "who did I pick, and what scoreline?", costing a three-tap trip to the bracket, repeated several times a day. Then a player asked to see leaguemates’ picks the same way.

2. ↓→

   Shipped

   A "Your picks today" digest at the top of the dashboard, then leaguemate pick grids and tap-to-compare, two days later.

3. ↓→

   Result

   114 players used the digest (4,769 views); league standings stayed the most-fired event in the app.

As shipped

Daily Kick

03

1. Signal

   Late joiners faced a 15-day dead zone between the lock and the Second Chance bracket, with nothing to do daily.

2. ↓→

   Shipped

   A one-tap daily poll with an animated global and league reveal, plus public archive pages. Works for every segment: leaders, busted brackets, and brand-new arrivals.

3. ↓→

   Result

   299 votes from 81 players across the run, on top of its SEO surface.

As shipped

Second Chance

04

1. Signal

   A real cohort arrived after the lock with no bracket to play, and busted veterans needed a reason to stay.

2. ↓→

   Shipped

   A knockout-only bracket seeded from the real 32 qualifiers: one perfectly fair board where late joiners and veterans pick with identical information.

3. ↓→

   Result

   101 players entered, and the app stayed alive for its newest users.

The network effect, by construction

The social loop was designed, not hoped for. Every new account auto-joins a global league, so nobody ever sees an empty leaderboard. Invite links unfurl into a league preview and complete the join after sign-in. The result: 80% of players joined a friend league (median size 5, largest 27), and 7 in 10 players arrived in a league someone else created. The product’s best growth channel was a friend with a bracket.

Built at a desk, played in a pocket

My raw analytics read roughly half desktop, but that desktop share was mostly me: filter out my own operating sessions and dev traffic, and players were three-quarters mobile before kickoff and 86% mobile across 32,500 tournament pageviews. The desktop experience I polished at my desk was, in effect, built for an audience of one. The tip jar taught the lesson best: it lived in a desktop sidebar, so most players never saw it. A mobile-visible tip tile shipped mid-tournament, and design for the pocket first became the standing rule for every surface that followed.

Spain beat Argentina 1-0 on July 19. Of the 146 complete brackets, 43 had called the champion from day one, a 29.5% hit rate, while 60.3% watched their pick go out before the Final. The global leaderboard was decided by 55 points, 756 to 701.

The tournament’s emotional peak is a 24-hour window, so the send had to land that night. "Footy26 VAR: Tournament Review", the per-player season recap covering placement, accuracy, round-by-round scoring, and bonus verdicts, went from approved spec to production sends in about 24 hours with zero new database migrations. Six pull requests shipped on Final day itself, and the thank-you edition reached all 177 reachable bracket builders that evening. Players who reviewed the season rated it 4.5 out of 5.

The sharpest thing the season taught me: I did not build a World Cup app. I built a social-prediction platform and shipped a World Cup app as its first tenant. The auth, leagues, invites, scoring rails, email machine, observability, design tokens, and 12-language chassis are sport-agnostic, and the privacy model was deliberately built so a second competition cannot cut corners: shared data stays counts-not-values, and every new competition must declare its own visibility rule.

The players saw it too: every season-review respondent independently asked for tennis and NBA brackets next. The architecture work is specced, with rolling per-round lock windows replacing the single global lock so a season stays alive for late joiners, and the 2027 Women’s World Cup, same sport, same flags, same audience, is close to a configuration change. The bracket is the beginning.