The story of Cinema.org

00 · At a glance

One question, 5,000 films

Cinema.org started with a simple frustration: I spend more time scrolling streaming menus than watching anything. Every service recommends what it owns, review sites tell you what’s popular, and none of them know my taste or tell me why a film is worth two hours of my evening.

So in early October 2026 I built the site I wanted: a curated catalog, a detailed taste survey, a plain-English reason for every pick, and an honest note on where to watch. I set the direction, made the calls and tested every version; Claude, Anthropic’s AI model, did the research, writing and engineering in Cowork. At the end of the month the catalog went from 3,500 films to 5,000.

• 956survey options across 35 questions
• 5,000films in the catalog
• 16themes, from heists to the inner life
• 1,414subtitled films, 28.3% of the catalog
• 7steps from a first list to 5,000

01 · The starting point

It began as a survey, not a website

The first thing I asked for was a way to describe my own taste properly: a “very very detailed” survey, useful for recommending things on Apple TV, Netflix and long-form YouTube. It became 35 questions with 956 options in eight parts: the basics, taste, range, practical habits, films seen, series seen, people and YouTube. Genres, countries, directors and actors get a three-way rating (love it, into it, not for me), a dealbreakers list covers things like heavy gore and shaky cam, and the “films you’ve seen” lists double as a never-recommend list.

Filling it in was the most useful hour of the project. I picked 235 answers: ten moods from “comfort watch” to “make me think for days”, subgenres like heists, time loops and cults, fifteen directors from Kubrick to Linklater, and all ten dealbreakers. One answer, subtitles “only if it’s great”, ended up shaping the whole catalog. The first thing we made from it was a profile page: 144 films I’d seen, 37 I loved, my dealbreakers, and 100 films I hadn’t seen, each with a one-line reason. Borderline content was flagged, not hidden. Looking back, that page already had the shape of the site: a profile that says what you love and avoid, and a ranked list where every film comes with a reason.

02 · How I asked

Short prompts, one decision at a time

My prompts were short, often dictated, and reacted to whatever was on the screen. I rarely wrote a spec:

• “ok now in a format like this, a self-contained html file”
• “can we now expand this to top250?”
• “lets take it to top1000 by adding 250 more english speaking films, and the 250 highest critically acclaimed films where they are not in english”
• “but it should be 5000, please add a few more, of any type, by the common directors in the rest”

Screenshots did the rest. Three screenshots and “not with these colors / layout” set the look; dictated percentages of gray became the dark palette. Claude asked when a choice was mine: at 1,000 films it pointed out my split would make the catalog a third subtitled, and I chose 200 subtitled plus 300 English instead.

03 · The catalog

From 100 films to 5,000

The catalog grew in seven steps, each with its own job. The first 100 came straight from my survey, sorted into 16 themes that describe how I actually choose, from Heists & cons to Identity & the inner life, plus Wildcards a step outside my comfort zone. The next 150 made a Top 250, screened against films I’d seen and my dealbreakers.

Films 251 to 1,000 turned a personal list into a real catalog: 126 canon films the survey named, pre-1970 classics, world cinema ranked across Sight & Sound, They Shoot Pictures and BBC Culture’s polls, and recent festival winners. Hard films like City of God stayed in with a content note instead of being dropped. Films 1,001 to 3,500 scaled up by scoring candidates half on critical standing and half on fit with my profile, then filling gaps: more languages, more silents, more recent releases. The last 1,500 went deep instead of wide, with a documentary shelf and more films by the catalog’s most frequent directors.

04 · Research

How every film was researched and written

Every film carries three kinds of writing. The description is three sentences: the premise with its leads, one specific true detail, and why it suits my taste. “Why it matters” is a neutral, spoiler-free note on awards, craft and influence. Tags, six to nine per film, start with the decade and come from one shared vocabulary.

Choosing films and writing them were separate jobs with separate briefs. The shortlist brief asked for real, well-regarded films the helper was certain existed, spread from the 1920s to the 2020s, each with its language, an eight-word reason and a check against every title already in the catalog. The writing was split into batches of 50 for parallel helpers, each given the same brief, my taste profile and examples of the house voice. The writer brief set hard limits:

• Description: 330 to 480 characters, ending “For your love of” and a taste hook.
• Why it matters: 220 to 340 characters, facts only.
• Style: American spelling, no dashes, and banned words such as masterpiece, iconic, gem and haunting.
• Tags: the decade first, then genre, mood, style and subject, in the vocabulary’s exact spellings.
• Content notes: only from a fixed list, such as animal harm, heavy gore or shaky cam.
• Accuracy first: never invent cast, characters or awards, and leave out any award you’re unsure of.

A script then checked every batch: titles, themes, tags, lengths, banned words and duplicates. One round of 1,000 films needed 84 fixes, mostly “why it matters” notes that ran short and descriptions that ran long. A separate fact-check of 91 sampled entries found four small errors, all fixed, such as the running time given for Russian Ark’s single shot. Streaming was the hardest part: a search cap stopped manual checks at about 150 films, so today 197 films have a listing and every film links to JustWatch for a live check.

05 · The documentary shelf

Films 3,501 to 5,000, step by step

The brief was to add the best documentaries the catalog didn’t have: the most informative, with the highest production value or acclaim. The catalog held 291, mostly from the last two decades, so the search aimed at older films and regions outside the US and Europe.

1. 1890s1 · 0.1%
2. 1910s1 · 0.1%
3. 1920s13 · 0.9%
4. 1930s16 · 1.1%
5. 1940s22 · 1.5%
6. 1950s28 · 1.9%
7. 1960s56 · 3.8%
8. 1970s85 · 5.8%
9. 1980s101 · 6.9%
10. 1990s123 · 8.4%
11. 2000s293 · 20.0%
12. 2010s509 · 34.8%
13. 2020s214 · 14.6%

Original 3,500Added

Documentaries weren’t scored against my taste; they were ranked by strength. Searches looked like these:

Academy Award for Best Documentary Feature Film nominees winners listSundance Grand Jury Prize Documentary winners list all yearsSight and Sound 50 greatest documentaries of all time poll full listYamagata International Documentary Film Festival Grand Prize winners list

Each source got its own search for one reference page holding the whole list, such as the Oscars or the Sight and Sound top 50, which was then read in full. Lists that couldn’t be reached were skipped, not guessed, and four lists I supplied myself (Criterion, A24, Kanopy and Venice Classics) were checked line by line. Every film went into one pool, duplicates were merged by title and year, and the pool was ranked: prizes first, wins ahead of nominations, then canon lists. After each pass the thinnest regions and decades got the next search: Asian festivals, Peabody winners, regional lists and pre-1970 films. By the last passes a third of the titles tried were already in the catalog, a sign of how complete it had become. 42 titles were cut on review, such as hybrids, TV series and unconfirmed winners, leaving 1,171 documentaries.

The last 329 places went to missing films by the catalog’s most frequent directors, led by Alfred Hitchcock (17), Ridley Scott (16) and John Ford (15). Every new film got a theme, region, tags and five related films, and the Explore cloud grew from 50 to 60 tags. 250 documentaries and all 329 director films have full write-ups; the other 921 documentaries show a placeholder for now.

06 · The market

Working out what was missing

There was no formal market study. The thinking came from being the first user, a strategy plan from a friend, and a comparison with the tools people already use. Every feature had to answer one question: does it get me to press play on something good tonight?

The friend’s plan gave the pitch, “What do you want to watch?”, and a measure of success: a film found and actually watched, not time on the site. I took its top five ideas, from natural-language search to a watchlist, and kept Google sign-in and an invite-only homepage against its advice: every member gets three invites, because good taste travels through people. That was a positioning call, made knowingly. The pitch is narrow on purpose: rankings can’t be bought, and as a 20% Public Good Organization, 20% of profits fund educational documentaries for underserved communities. That pledge still has to match our governing documents before launch.

07 · Design

Minimal, dark and accessible by default

The site uses the MINML gray style: DM Sans for text, DM Mono for labels and one green accent for anything selected. The dark theme came from a dictated note of gray percentages. The build checks 26 color pairs for 4.5:1 contrast, keeps touch targets at 44 pixels, limits motion to a short fade, and makes every first screen readable with scripts off, which is how a Mac previews a file.

08 · The product

What the site does, version by version

• Beta v1 to v3: the survey and a ranked list in one app, then a mood-aware search box, picks after five answers, why-it-matters notes and a watchlist.
• v4: a 0 to 100 score on every film, a cinema map, seven taste sliders, invites and Explore by mood.
• v5 to v13: 1,000, then 2,500, then 3,500 films; Explore became the home page with up to three tags, and every list was reworked for phones.
• 5,000 films: a 60-tag cloud, About, Analysis and Story rebuilt on the whole catalog, All films back with 120 tags, and a nomination button on every page.

The ranking is plain arithmetic. Films score points for matching your moods, genres, directors, eras and countries, and lose points for what you avoid; a content note that hits a dealbreaker costs eight points. Each pick shows its top reasons, like “Like Heat, which you loved”.

09 · Engineering

How it was built, and what broke

For most of the build every page was one self-contained file, which made Explore a 9 MB page at 5,000 films. The site is now a small folder: light pages, shared fonts and styles in assets/, and film data in data/, loaded as needed, so Explore’s first screen appears in a fraction of a second. Marks stay in your browser; there’s no server yet. Scripts merge the data, run the contrast and syntax checks, and test every page and link.

• The sandbox filled up mid-build, and we finished in a fresh session from a zip.
• One line broke every link when the folder was opened from disk. Every link is now checked by script.
• The Mac preview was blank because the first screen waited on an animation. It no longer does.
• The first draft of this history was wrong until the saved snapshots set it straight.

The lesson: keep the build folder portable, zip it after every milestone, and make every check automatic.

10 · Where it stands

Before and after

Where the additions went, by region

1. United States2,514 · 50.3%
2. Europe930 · 18.6%
3. UK & Ireland555 · 11.1%
4. East & Southeast Asia423 · 8.5%
5. Middle East & Africa195 · 3.9%
6. Latin America188 · 3.8%
7. Oceania104 · 2.1%
8. South Asia91 · 1.8%

Original 3,500Added

Still being finished

• 921 of the 1,171 added documentaries show a placeholder instead of a full write-up.
• The added write-ups are drafts and will be rewritten to the house brief.
• 996 added titles still need their title, year and director verified.
• Streaming has been checked for 197 films; the rest say so plainly.
• Watched, loved and liked numbers on film cards are demo numbers.

11 · What’s next

From a beta for one to a site for many

• Streaming for all 5,000 films, from TMDB’s watch-provider data, refreshed weekly.
• Every added film checked and written to the house brief.
• Real members, with proper sign-in, a database and a privacy policy, invited in small waves.
• Smarter search that understands “something warm and strange for a rainy Sunday”.

The promise stays the same: one decision, made well. Find something extraordinary tonight, know why it’s worth your time, and see where to watch it.

Developed by G with Claude (Anthropic) in Cowork · October 2026

Explore the filmsBrowse all filmsAbout Cinema.orgSee the analysis