↗ expand01 / Entry
Two modes, not one archive
New users choose quick relief or taste-building.
Independent product case · live MVP
From random picker to taste-aware discovery.
Cine.pick started as a weird little movie randomizer. Fun once, not enough to come back to. I rebuilt it into a mood-first discovery MVP with two paths: quick help when you are tired, and taste profiling when you want to understand what you actually like.
This is an independent project, not a commercial product. I use it to show product thinking, interaction design, visual craft, front-end implementation, and where I draw the line between what shipped and what still needs validation.
Product framing
Consumer UX
Mobile-first flows
Design system
AI-assisted build
Live MVP
6weeksrebuild sprint
54filmscurated database
2entry modesdifferent user states
5interviewsdirectional research
LiveMVPnot just Figma
Product walkthrough
First, see the loop.
Mood input → quick recommendation → Movie DNA → taste profile.
02
The project did not start clever.
The first Cine.pick was closer to a playful button than a product: choose a number, get random films. I liked the energy, but the product had no memory. No reason to return. No sense that it understood you.
A randomizer with attitude
The early visual direction was loud, acid, almost Spotify-ish. It solved one tiny moment: “just give me something”.
01Pick a number of films
02Get random results
03Maybe watch one, maybe leave
From one-off fun to a product loop
The better question was not “can I randomize films?” It was “can the product learn enough about taste to make the next choice easier?”
The tired watcher
“Please decide for me.”
Low energy, no filters, no scrolling. Needs a result fast and a short reason to trust it.
The visual curator
“Show me what my taste says.”
Wants aesthetic discovery, taste patterns, deep search, and a profile that feels personal.
03
Research scope: useful, not overclaimed.
I did not run market validation. I used directional interviews and prototype feedback to find where the randomizer idea broke, then changed the product loop around those signals.
Was “random” actually helpful?
- Did random choice feel like relief or a gimmick?
- When did users want fast help versus deeper exploration?
- Did people need an explanation before trusting a recommendation?
People had different decision energy.
- Some wanted the product to decide quickly.
- Some wanted to build a taste profile and understand patterns.
- “Why this film?” mattered more than another beautiful card.
I changed structure, not decoration.
- Archive-first onboarding became two entry modes.
- Movie DNA became the trust layer.
- Missing data states became part of the MVP instead of a later polish task.
Honest limit: 5 interviews is discovery-phase scope, not validation scope. This helped shape a coherent MVP, but it does not prove retention, recommendation accuracy, market fit or monetization. Before shipping this as a real product, I would run a 15–20 participant task-based round around the bifurcated onboarding, recommendation trust, and guest-to-registration moment.
04
The key redesign: stop sending new users to the archive.
“Explore the Archive” sounded useful, but it pushed people into browsing. That was exactly the behavior I wanted to reduce. I replaced it with two action-first paths.
Before
Explore the Archive
A library-first CTA duplicated the top navigation and did not solve cold start.
After
Analyze my DNA / No brain cells tonight
The onboarding now asks what kind of help the user needs: build taste or get a quick recommendation.
05
Three moments had to work.
I kept the final story small: choose a mode, get a recommendation, understand why it fits.
Quick relief
The Void decides
For tired users, the product should reduce choices instead of asking for more input.
Short test
Enough signal, not a quiz maze
A few questions can create a more interesting result without becoming the whole product.
Taste building
Analyze my DNA
Swiping solves cold start faster than asking a new user to rebuild their whole film history.
06
Testing signals → design decisions.
No fake metrics here. These are qualitative signals from interviews and prototype walkthroughs: what changed, what went to backlog, and what I consciously accepted as a trade-off.
Split the entry by user state
SignalPeople arrived with different energy: some wanted exploration, some wanted relief.
DecisionStop treating every new user as someone ready to browse.
ChangeTwo CTAs instead of one archive path.
Made the DNA teaser legible
SignalThe old teaser looked atmospheric but did not explain why sign-up mattered.
DecisionShow value before asking for commitment.
ChangeClearer profile preview, demo link, and visible taste reward.
Designed for missing content
SignalA small database means some film pages will not have rich stills or deep dive data.
DecisionDo not let thin data look like a broken product.
ChangeEmpty states and same-vibe collections keep the page useful.
“There is only one film in Editor’s Pick — it would be cool to make a carousel.”
The ask was valid. A single hero film was an intentional MVP constraint: curation signal over choice overload. But it does limit repeat visits. I put carousel exploration in the backlog for when the content library has enough editorial depth; adding it too early would create empty states worse than the single-hero version.
“Add to Watchlist could change colour or animate to confirm.”
Missed state. I had designed the trigger, but not the confirmation. That was a happy-path gap. I deferred it out of the first MVP build, but documented it as a P1 polish item: added → confirmed state, with a small haptic-style micro-interaction.
Some feedback confirmed the direction without changing the design.
“There is no service that offers movies based on mood.”Supported the mood-first direction as a useful niche, without proving market demand.
“I want psychological subtext in the choice.”Supported Movie DNA as an explanation layer, not just a decorative profile.
“Ten films in CineMatch is quite a good number.”Supported keeping the cold-start swipe flow short instead of asking for a long taste setup.
How I use this: validated findings are weaker than changes. They justify continuing the direction; they do not prove it is right. The harder work is still in the iterations and next validation plan.
“I wanted to register, but I couldn’t.”
Conscious decision. Registration is disabled in the demo because Cine.pick is a portfolio MVP, not a live commercial product. The cost is real: some people ignore the entry disclaimer, hit a wall, and lose immersion. I accepted that for this build because real auth was not worth the backend scope. In a live product, registration would become the main conversion target of the CineMatch guest flow, and the DNA teaser is designed for that future moment.
07
The product model became simpler.
The MVP does not need every social feature, review system, or streaming integration. It needs one understandable loop.
IntentMood, tiredness, curiosity, or a specific aesthetic craving.
SignalRandomizer choice, short test answer, swipe, like, dislike, watchlist.
ExplanationMovie DNA shows themes, tone, pace, and why the match makes sense.
MemoryThe product gets better as the user reacts, saves, or corrects tags.
I did not build a social review feed.
Open reviews would add noise and moderation debt. If Cine.pick ever becomes social, I would start with structured reactions: agree/disagree with vibe tags, suggest a tag from a controlled list, and taste compatibility with friends.
I did not make “where to watch” the main point.
The product is about discovery and taste, not only consumption. Streaming links can help, but they should not flatten the experience into a utility page.
I kept the editorial feel on purpose.
The dark, cinematic UI is not just trend styling. It supports the idea of a niche film journal with product logic underneath.
08
Small edge cases, less broken magic.
A recommendation product loses trust quickly when pages feel empty. I added a few small states so the MVP does not collapse when the database is thin.
implemented
No film stills yet
The gallery does not pretend data exists. It says stills are coming and keeps the layout calm.
implemented pattern
No deep dive data
Instead of leaving the editorial section empty, the page can point to films with a similar vibe.
melancholic
agreenot quite
next
Tag feedback
Agree / disagree with a mood tag would help the database learn without opening the door to messy open-ended UGC.
09
AI helped me move faster. It did not decide the product.
I used AI like a build partner: useful for scaffolding, debugging, and variants. The product choices stayed mine.
AI helped with
01
Implementation scaffoldingPHP, JS, CSS structure and small interaction states.
02
Debugging and edge casesLayout issues, empty states, small database logic.
03
Content and tag explorationDrafting possible moods, vibe tags, and film metadata patterns.
I owned
01
Problem framingFrom random result to taste-aware discovery.
02
Interaction modelTwo entry modes, Movie DNA, cold start, and scope cuts.
03
Final judgmentVisual system, QA, what to ship, and what not to claim yet.
10
Outcome and what I would test next
This is a working MVP ready for structured validation, not a proven market product. That distinction matters.
What shipped
- Live responsive MVP
- Two onboarding paths
- Randomizer and short mood test
- CineMatch swipe flow
- Movie DNA profile
- Edge states for missing content
What I would measure next
- Time to first useful recommendation
- Onboarding comprehension
- Perceived fit of the result
- Save / watch intent
- Guest-to-registration intent
- Return sessions
What I would not claim yet
- Market fit
- Recommendation accuracy at scale
- Retention
- Monetization
- Social behavior
- Large catalogue quality
Reflection
The useful move was not making the app smarter. It was making the first decision smaller.
What this case says about my work
I like products with a strong mood, but I do not want the mood to hide weak logic. This case helped me practice cutting the cute parts, keeping the useful parts, and showing the limits clearly.
The next version would need better recommendation data, a cleaner DNA model, and a proper validation study. I know that. That is the point.