B
1. The rules, read before anything else
Devpost is strict about the video, so the rules were copied literally into the first lines of the script file on September 12: under two minutes, public on YouTube or Vimeo, it must show the app working on the real device, no third-party music or trademarks, in English or with English subtitles. After September 30 it cannot be edited.
Two more rules were mine. Nothing in the video promises what the app does not do. And nothing inside a screenshot gets retouched. Both of them cost work later. Both were worth it.
2. The script came before the camera
I dictated three blocks in this order: the emotional part, the design, the technique. The order stayed. What changed was the shape.
The judges watch dozens of videos, so the first four seconds carry one sentence in my voice: "Being me was always a battle." The bullying is one line, not a story. The video is framed by training, not by damage. It opens with the battle and closes with "I trained for years to stop being afraid of being seen. I built AURADUEL so you don't have to wait that long."
The script does not name why I was bullied. That is a personal decision, and the text works without it.
Before I saw it, the script went through two of my AI reviewers. The behavioral one scores every piece against a rubric and needs a 7 to pass: it gave 47 out of 60, a 7.83. The brand guardian took five passes to approve. One sentence it would never let through is "ranks climb with every battle", because that is not what the code does. The script says "you earn your rank battle by battle".
3. The decision: no bots in the stands
AURADUEL has a crowd mode. Spectators scan a QR, enter the stands from their own phones, send reactions and vote. I was filming with two people and no audience. The obvious trick was on the table, because the tool already exists: a test harness that fills a room with simulated voters.
The answer was no, for four reasons that are written in the script file.
- The project's own anti-scam law forbids it. The header of that harness says it is only for my test rounds and never for public material. The HUD is honest by design: simulated spectators are labelled SIMULATED ROOM, real ones are labelled CROWD. A fake vote without the label deceives the jury. With the label, it confuses them.
- The rules ask for footage of the project functioning. Bot votes are not the product functioning. If a judge tries the stands with a friend and does not see what the video showed, I lose credibility exactly where I am asking for it.
- It is the wrong symbol. One of the hard rules of the game is "no vote buying". Showing bots voting in the demo of an app that competes on integrity contradicts the product inside its own video.
- It costs hours and adds nothing. The show was already there: a real duel with the aura drawn on top.
Opening three browsers ourselves to fake a bigger crowd was discarded too. It is the same trick with more steps.
What the video shows instead is the real door: the lobby with the real QR, and one real spectator on a third iPhone. The result screen prints exactly what happened: crowd vote 1 to 0, one person in the stands. The measured score decided that round. That is what the video shows: the real door, not a staged result. The subtitle says friends can scan a QR and vote. It does not say a crowd did.
4. Shooting day: two people, three iPhones
September 13, by daylight. The field QA notes of the project mark night as the hardest condition for body detection, and the design rule "dark first" is about the interface, not about the footage.
The golden rule of the day: the app runs on one iPhone and another device films it. We recorded five takes. One solo round of "Measure your aura", 31 seconds. Then one mirror duel between my partner and me, captured four ways at once: the screen recording of each player's phone, a fixed aerial shot of the two of us with a phone on a tripod in front of each, and the third iPhone entering the stands by QR and voting.
Only two game modes were allowed on camera: solo and mirror. The local referee mode starts ambient bots, and the HUD would have printed SIMULATED ROOM in the middle of my demo. The checklist has a line for that: zero frames with that label. My partner signed an image release, one page. Nobody else appears, so there were no minors and no third parties to manage.
The store screens were recorded with the app switched to English: the wardrobe, a cosmetic at $2.99, the Apple payment sheet in sandbox, and the Premium screen with its sentence, "Premium never improves your result. Not one aura point, not one advantage: that's earned by playing." One of the test accounts is on the Swiss store, so its Apple sheet came out in German with prices in francs. That one never made the cut.
5. Frame by frame
I cannot watch five videos and remember where everything is. So Claude did what an assistant editor does. It extracted one frame every half second from each screen recording, one per second from the aerial shot, and zoomed at ten frames per second into the rings and the corner where crowd reactions appear. The result is a table per take: second by second, what is on screen.
That density was a correction of mine. The first pass sampled one frame every two seconds and declared that some shots were missing. They were not. The final result screen lasts 1.9 seconds, the consent screen 0.8, the tap on the vote 0.4. Nobody gets to say "it is not in the footage" after looking at one frame every two seconds.
That table found things I had not seen. Both screen recordings have internal cuts. The aerial shot has no audio. There is a
reaction from the stands floating up the corner of one phone for half a second. And there are exactly three windows where the sources can be synchronized, using the countdown on the phones and my high kicks as the clapperboard. The best one lasts 4.5 seconds: the aerial shot and both phones at the same instant, one screen saying "Round for you" and the other "Round for your rival".The same analysis produced a list called "never in the cut": sandbox purchase errors, the iOS screen-capture UI, the German payment sheet, an "update available" toast, and the line of the Apple sheet that shows an account email. The red recording pill of iOS is removed by framing, never by painting over it.
6. English without touching a single screenshot
We recorded the game in Spanish. It is built for the Latino community first, and that was my decision. But the jury reads English.
The easy way is to edit the screenshots. I did not want a single pixel changed inside the phone, because the video has to show the project functioning. So the labels live outside the capture: an English word on the dark background, in the same typeface the app uses, with a thin cyan line that goes to the edge of the screenshot at the exact height of the Spanish word. It appears and disappears with the word. It works like a live concept map.
The part I like most: every label must be the official English string of the app, taken from its translation files. The script that renders the labels has a verify flag that checks each text exists literally in those files, and fails if it does not. The only exceptions are three words from Apple's payment sheet, marked as our own translation. So the English in the video is not a subtitle I invented. It is what the app itself says when you switch the language.
The first version of that tool had a bug you could see: with some alignments, the line crossed out its own label. Now the line is traced after the text is laid out, and it leaves from the side that faces the capture.
7. A rough cut with cards, then my voice, then one evening
At 1:18 in the morning of September 14 there was a v0: the full structure, with placeholder cards where my face would go, so I could see durations before recording anything. That evening I recorded the whole speech in Spanish at home, 113.9 seconds.
The final cut was assembled that same night by a separate Claude session working from a written brief. My voice drives the timeline: it is transcribed with word-level timestamps, and each screen enters when I name the thing. My face is never covered. I stay on the left third and the phones enter on the right. By 23:03 there was a v1.
The transcription also caught what I did wrong.
I skipped about 60 words of scene 7, the one that explains privacy, the QR and the vote. There was only one take. With the full text the video would have passed 130 seconds. The decision was mine: a 5.4-second stretch without my face, on a light background, with the stands on a phone and an English concept map, and one sentence of voice-over that I recorded afterwards with the same microphone. What did not fit went into the written submission for the judges.
I said the wrong word in scene 6. The rule is "the interface never competes with the person". I said "the battle never competes against the person". I re-recorded the sentence, take two of two, and it plays over a shot where the person is in the center and the HUD stays at the edges.
Three ideas came from me in the heat of the edit: the opening, where the logo orb charges, bursts, and the burst is the mask that reveals my face centered in the ring; the light stretch without me; and the music, a track of my own that I made with Gemini, mixed very low under the voice.
8. Measured, not assumed
The brief ended with a checklist, and a script runs it against the finished file. These are its numbers for v1:
- Duration: 117.40 seconds, against an internal cap of 118.
- Loudness of the final mix: −14.1 LUFS.
- Black frames: 0. Forbidden segments used: 0.
- Nothing covers my face in any of my shots.
- Vetoed words in our own texts and subtitles: 0.
- Sync: the final mix is transcribed again, and 30 of 31 subtitles start within a third of a second of their word. The other one reads late only because the lyrics of the song confuse the transcriber. Measured on the voice alone, it is a quarter of a second.
9. The guardians said no to me, too
The brand guardian approved the first pass with four minor notes. Then I changed the music, and it vetoed three verses of my own song. The lyrics implied things the game does not do: physical impact, and a crown promised in advance. The fix was one jump in the music bed, placed between complete phrases, that leaves two of the verses out. The third one stayed by my decision, as a metaphor, and that decision is written down with my name on it. The third pass approved.
A code review of the editing scripts found something better than a bug: a check that could never fail. The opening and closing durations were typed by hand in two places, so the frame count was always "right". Now they come from one edit list and the exact number is required.
And the light background of that 5.4-second stretch is not in the brand palette. It is recorded as an exception, limited to that stretch, with a soft fade so there is no flash.
10. What I learned
- Read the contest rules first and paste them on top of the script. They decide the structure.
- If your product is about fair play, the demo has to play fair. Show the real door, not a staged result.
- Do not retouch screenshots. Explain them from the outside, with the strings your app already ships.
- Let your voice drive the edit. Transcribe it before cutting: you did not say what you think you said.
- A rough cut with placeholder cards saves you from recording things twice.
- Measure the final file: duration, loudness, black frames, sync. "It looks fine" is not a measurement.
- Record who decided what. When a guardian says no and you overrule it, sign it.
The video is 117 seconds long. From the first script to the cut I approved it took three days, September 12 to 14: a weekend and a Monday night. It is public on YouTube, and the app it shows is the one you can download.
Built for RevenueCat Shipaton 2026. AURADUEL is on the App Store
Demo video: