Personal project / 03

Never miss a
new season.

Track your favourite shows and get notified the moment a new season is confirmed, dated, or ready to stream, and exactly where to watch it.

Role
Product & UX direction
Build partner
Claude Code
Timeline
3 weeks, self-directed
Status
Working prototype: real auth, data & notifications

Status system: live preview

● Renewed ● Now airing ● On hiatus ● Rumored ● Ended

Five states, each with its own icon and colour, not colour alone, so status reads correctly at a glance for every user, including those who can't distinguish it by hue.

The problem

Following a show shouldn't mean checking five apps.

This started from my own habit, not a formal research study. Shows are scattered across platforms, "renewed" news is usually just a rumour repeated across fan sites, and there's no single reliable place that tells you: is this actually confirmed, when's it landing, and where do I watch it. It didn't stay a hunch for long; see below.

Signal problem

Renewal news is noisy: rumour, leak and official confirmation all look identical scrolling a timeline. Users need those told apart, not blended together.

Trust problem

Where-to-watch information goes stale fast as licensing deals shift. A tracker that's wrong erodes trust faster than one that's simply incomplete.

Attention problem

A tracker only works if updates reach you: checking an app is a habit most people won't form on their own, which turned out to matter more than expected (see the feedback pivot below).

Research

What the data, and other people's reviews, already say about this.

Not primary user interviews. Desk research: published survey data on streaming behaviour, plus real reviews of the trackers people already use to solve this exact problem. Three findings kept coming up, and each one maps to a specific decision in Next Season.

4
average paid streaming subscriptions per US household; 90% of households now have at least one

Source: Deloitte, 2025 Digital Media Trends

110 hrs
a year the average American spends just deciding what to watch: nearly five full days

Source: UserTesting / Talker Research

1 in 5
viewers give up finding something to watch and abandon the session with nothing on

Source: Nielsen, State of Play

What the trackers people already use get wrong

"Wrong where-to-watch, often"

JustWatch reviewers report "where to watch" data being wrong often enough that some stop trusting it outright: reviews cite specific seasons still listed on platforms that dropped them long ago.

Source: JustWatch reviews, Trustpilot

→ Fix: a visible "Checked Xh ago" timestamp on every card, synced straight from TMDB rather than an opaque aggregator, so the source is shown, not just asserted.

"Still shows last year's season"

TV Time reviewers describe the app still listing a show at last year's season well after a new one aired: status information that's gone stale doesn't just fail to help, it actively misleads.

Source: TV Time reviews

→ Fix: five explicit statuses (renewed / airing / hiatus / rumored / ended) refreshed from live data, instead of one label that quietly falls out of date.

"It gets lost at the bottom of the list and I miss it"

A Trakt reviewer's exact complaint: without a way to surface what just changed, a new season landing looks no different from nothing happening. You only find out if you go looking for it.

Source: Trakt reviews

→ Fix: a dedicated Recent Updates feed first, then, once testing showed pull-only updates weren't enough, real push notifications by email.

Who this is for

Proto-personas built from the desk research above, not from primary interviews: there was no user testing on this project, so these describe behaviour patterns evidenced in the data and reviews, not real people I spoke to. Each one maps to one of the three problems above.

MC

The Multi-Platform Completionist

4+ services, 8-10 shows tracked

Has been burned by a tracker listing a show on a platform that dropped it months ago, so now double-checks everything manually, which defeats the point of using a tracker at all.

"I don't trust the app's answer until I've Googled it anyway."

Signal problemTrust problem
DF

The Decision-Fatigued Scroller

Spends real time just deciding what to watch

Wants to know instantly whether "renewed" is real news or fan-site rumour, so they're not chasing a show that isn't actually coming back.

"Half the 'Season 5 confirmed' posts I see turn out to be someone's wishlist."

Signal problem
HN

The Habitual Non-Checker

Opens the tracker weekly, not daily

Wants to be told when something changes rather than remembering to go looking. A new season landing looks identical to nothing happening unless they check that exact week.

"If it's not in my inbox, it didn't happen as far as I'm concerned."

Attention problem

How this was built

Directing product & UX, with Claude Code as build partner.

What "I built this" means here

I directed every product and UX decision (what to build, why, and exactly how it should look, feel and behave), then worked with Claude Code as my build partner to turn those decisions into real, working code rather than static mockups. That's a genuinely different loop from designing screens in Figma: every decision gets pressure-tested against real, running software the same day.

My job in that loop was direction, review and judgement: specifying what "confirmed" vs "rumoured" should mean and how to signal it, calling the accessibility audit and reviewing its fixes, catching the bugs that only showed up once real data replaced my mock data (see Resilience), and reprioritising the roadmap when testing exposed that updates were easy to miss.

Process

Four phases, each one closing a gap the last one exposed.

Built across three weeks of sessions rather than one continuous sprint. Each phase started by using what existed, finding what was wrong with it, and fixing that specifically rather than moving on.

31 Jul

Scaffold & shape
Core watchlist loop on mock data, then a visual pass: ambient glow, glass nav, stats strip, a discover row so the page has content from day one.

3 Aug

Depth & trust
Countdown dates, an "on hiatus" status distinct from airing/ended, a "Checked Xh ago" trust signal, genre-weighted recommendations.

4 Aug

Accessibility & real data
Full WCAG AA contrast audit and fixes, then swapped mock data for live TMDB data, and fixed what that broke.

18 Aug

Accounts & the pivot
Magic-link auth with cross-device sync, then closed the biggest gap testing exposed: real email notifications.

The product

A watchlist, a trust signal, and a way back in.

Three screens carry the whole loop: track a show, see what changed, add more without losing your place.

Dashboard: watchlist, recent updates, recommendations

nextseason.app
◆ NEXT SEASON
+ Add show

Never miss a new season.

Track your favourite shows and we'll tell you the moment a new season is confirmed, dated, or ready to stream, and exactly where to watch it.

◆ 3 Tracked ◆ 2 Confirmed ◆ 0 Rumored
Stranger Things poster★ 8.6
Stranger Things
Netflix · Season 4
● Renewed

Season 5 confirmed, final season · Nov 26, 2026 · In 3 months

Checked 2h ago

House of the Dragon poster★ 8.4
House of the Dragon
HBO Max · Season 3
● On hiatus

On hiatus between seasons: no Season 4 date yet

Checked 5h ago

Severance poster★ 8.7
Severance
Apple TV+ · Season 2
● Renewed

Season 3 confirmed · Jan 15, 2027

Checked 1d ago

Because you're watching Severance

The Last of Us poster

The Last of Us

Wednesday poster

Wednesday

Fallout poster

Fallout

Add a show: searches the loaded catalog and TMDB live

nextseason.app
Add a show×
Search any show…
Fallout poster

Fallout

Prime Video

Wednesday poster

Wednesday

Netflix

The Sopranos poster

The Sopranos

HBO Max

Real dialog semantics, not a styled div: role="dialog", focus moves in on open and restores to the triggering button on close, and the page behind is marked inert while it's open, so a manual focus trap isn't needed, because inert content simply isn't reachable by Tab.

Sign in: passwordless, email magic-link

nextseason.app/auth/sign-in

Accessibility

An audit against numbers, not a squint test.

Rather than eyeball the "Cinematic Indigo" palette and hope, I computed the actual contrast ratio for every text/background pair against WCAG AA (4.5:1 for normal text, 3:1 for large text and UI components), including alpha-composited badge backgrounds. Four pairs failed. Here's the fix, with the real before/after ratios.

Muted text (counts, "Checked Xh ago") on background
#6a6e91 · 3.89:1
→
#8488a5 · 5.53:1
AA pass
White text on the primary button fill
#7c6cf6 · 3.94:1
→
#705ff5 · 4.52:1
AA pass
White text on the button's hover state
#9084f9 · 3.06:1
→
#4934f2 · 6.87:1
AA pass
"Ended" status badge text on its own soft background
#8b8f9c · 4.38:1
→
#8e929f · 4.50:1
AA pass

Darkened, not lightened, on hover

The button's hover state (#4934f2) is darker than its resting fill (#705ff5), not lighter. That's a deliberate choice so hovering improves contrast instead of regressing it, which a naive "lighten on hover" pattern would have done here.

Beyond colour

A keyboard bug the contrast audit surfaced along the way

The Add Show search field used focus:outline-none with nothing replacing it: invisible to sighted keyboard users, and to no one using a screen reader either, since there was simply no focus indicator to announce. Fixed with a visible accent-coloured focus ring.

Landmarks & structure

NavBar and main content use real <nav>/<main> landmarks instead of generic divs, so screen-reader users can jump straight to content.

Motion

The card fade-in animation respects prefers-reduced-motion and is skipped entirely for anyone who's set that preference.

Decorative vs. meaningful content

A poster's fallback initials are aria-hidden (redundant, confusing next to the real title read aloud), but the rating chip gets a real label, "Rating 8.6 out of 10", instead of a bare icon and number.

Dialog semantics

Add Show is a real dialog: role="dialog", aria-modal, a labelled search input, focus moved in on open and restored to the trigger on close.

Resilience

What real data broke that mock data hid.

An 11-show test catalog is forgiving. A live ~100-show TMDB feed is not: swapping mock data for the real thing surfaced two failures that never showed up once in testing against the mock set.

Bug: recommendations vanished entirely

The "Because you're watching X" rows used a fixed similarity floor (0.25), calibrated against the small test catalog. TMDB only has ~16 genre categories total, so spread across ~100 real shows almost every genre is shared by far more shows than that floor implicitly assumed. Result on real data: every recommendation row disappeared.

Fix: derive the floor from the catalog itself: the weight of whichever genre is most common sets the bar a match has to beat, scaling automatically whether the catalog has 10 shows or 500.

Design decision: fail partial, not total

This sandbox can't reach TMDB to verify the real platform's function timeout, so rather than risk an uncontrolled timeout returning nothing, the catalog fetch carries its own 8-second time budget and bails out gracefully, returning whatever real shows it already has rather than failing the whole page. If TMDB is unreachable at all, or no API key is set, the app falls back to the mock catalog silently rather than crashing.

Why this matters for a real service

Mock data is convenient precisely because it hides the failure modes real data has: uneven distributions, rate limits, timeouts, missing fields. Treating "works on my test data" as done is how services ship broken to production. The fix here wasn't cosmetic, it was catching that the feature had silently stopped working at all.

Iterating on feedback

The biggest gap wasn't a feature. It was that updates were pull-only.

Using the product for real, not just building it, surfaced the actual priority: a tracker that only tells you something when you remember to open it isn't solving the problem it exists to solve. That reprioritised the whole next phase.

Before

Updates only appeared in the dashboard's "Recent updates" feed, visible only if you opened the app after a change happened.

Built

Real accounts (magic-link auth, Supabase, row-level security) so a watchlist persists and syncs across devices: the precondition for reaching someone who isn't looking at the app right now.

Closed the loop

A daily Vercel Cron job reruns the exact same change-detection logic the dashboard uses, for every user, and emails anyone with a real change: a branded template via Resend and react-email, not a generic transactional email.

Architecture

Built like a real service, not a demo.

Data & privacy

Postgres + row-level security

Three tables (watchlist_items, show_snapshots, user_state), each scoped by Postgres row-level security so a signed-in user's queries can only ever touch their own rows. Privacy by construction, not by convention in application code.

Live data

TMDB, deduped & time-budgeted

Popular, Top Rated and On The Air lists merged and deduped (not one list paged deeper), a live search route for anything not pre-loaded, and the graceful partial-failure handling covered above.

Notifications

Cron + the same detection logic

The daily email job reuses the identical detectUpdates() function the dashboard renders from: one source of truth for "what changed," not a second implementation that can drift out of sync.

Visual system

Cinematic Indigo.

A blue-violet-tinted dark, the same instinct behind Prime/Disney+/HBO's dark UIs, tuned to its own violet accent rather than borrowing anyone else's. Every status colour was chosen to also carry its own icon and label, never colour alone.

Background

Surface

Accent

Renewed

Rumored

● Renewed ● Now airing ● On hiatus ● Rumored ● Ended

Space Grotesk: Headline

Never miss a new season.

Inter: Body

Set for a dashboard you'll glance at often, not read for long: high enough contrast to scan fast, quiet enough not to compete with poster art.

Reflection

What this demonstrates, and what's still not done.

  • An accessibility audit run against real computed contrast ratios and fixed with evidence, not a visual guess, including a keyboard bug that guessing would have missed entirely.
  • Resilience: the recommendation engine and the data fetch were both designed to degrade gracefully rather than fail completely when reality didn't match the plan.
  • Prioritisation driven by using the product for real, not by preference: the pivot to notifications happened because pull-only updates were the actual gap, not because it seemed like a nice feature.
  • End-to-end thinking beyond the UI: real auth, row-level-secured data, and a notification pipeline that shares its logic with the dashboard rather than duplicating it.

What I'd do next

Honest gaps

This is a working prototype, not a service with real users yet. The next steps are real usability testing, including with people who use assistive technology (the WCAG audit checks the numbers; it doesn't replace watching someone actually use it), JustWatch integration so "where to watch" data is definitive rather than TMDB's network field, and monitoring on the cron job so a silent notification failure is never the way a change gets missed.