aloud
eight ways of reading a video
It downloads the clip, reads every word on screen, transcribes the audio, checks the claims against the web, and measures how it did against everything else the creator has posted. Two to four minutes.
starterstory · 33.2s
- This is a short clip where a SaaS founder with nearly a million dollars in annual revenue lists every tool in his 2026 tech stack.
- He says Convex handles the back end and database, though on-screen text flashes both FREE and $40 a month for it.
- Vercel hosts the front end, and Railway hosts a separate microservice that builds the actual content.
- Clerk handles authentication, and Resend sends transactional and mailing list emails.
- Axiom is used for logs and observability, shown on screen at $25 a month.
- OpenAI and Claude models generate copy and captions for the product, listed as free.
- Fal.ai handles all image and video generation, with on-screen text pricing it at $5,000, far above every other tool listed.
- He closes by saying Claude Code on the 20x Max plan, shown as $200 a month, is essential for how fast his team ships.
- On-screen price tags appear throughout but are never read aloud, so the video conveys cost information purely visually.
fact checks
5 live web searches ran for these, and every source below was fetched to confirm it loads. Searched for: Convex backend database pricing free tier, Claude Code Max plan $200/month 20x usage, Resend transactional email API company, Axiom.co logs observability platform pricing, fal.ai image video generation API pricing.
“We use Convex for our back end and database”
supported Convex is a real backend-as-a-service platform offering database, auth, and serverless functions, with a free starter tier and paid plans around $25-40 per user per month.
“We use Claude code on the 20x Max plan”
supported Anthropic's Claude Max plan does have a 20x tier priced at $200/month, offering roughly 20 times the usage limits of the base Pro plan.
https://intuitionlabs.ai/articles/claude-max-plan-pricing-usage-limits
“We use Resend for all our transactional and mailing list emails”
supported Resend is a real developer-focused transactional email API company founded in 2023 and backed by Y Combinator.
“We use Axiom for logs and observability”
supported Axiom.co is a real logs, metrics, and observability platform with a free personal tier and a paid Cloud plan starting at $25/month.
“We use FoulAI (fal.ai) for all our image and video generation”
supported Fal.ai is a real serverless API platform offering over 1,000 image, video, and audio generation models billed on usage.
what it is doing
This is a short promotional video from starterstory in which a founder of a nearly million-dollar SaaS business lists out the tools that make up his company's tech stack.
The voice walks through each tool in the stack in order — Convex, Vercel, Railway, Clerk, Resend, Axiom, OpenAI and Claude, fal.ai, and Claude Code — narrating what each one is used for. The screen echoes almost every one of those names and functions as on-screen labels, but it also adds a layer the voice never touches: a monthly price and a URL next to most tools, for example $40/mo next to convex.dev, $150/mo near railway.com, $25/mo by clerk.com, $100/mo by resend.com, $25/mo by axiom.co, $5k near fal.ai, and $200/mo appearing at 30.0s. None of these figures are ever spoken aloud.
A listener who only listened would never learn the specific monthly costs or the $5k figure tied to the image-and-video generation section that the screen displays alongside the tool names.
shown on screen, never spoken
$5k is the only value on screen without “/mo” on it. If it is also /mo, the total is $5,540 and $5k alone is 90% of it. If not, $540/mo.
the first 3 seconds
the frame a second in
what you hear first
“What's the tech stack behind this nearly million dollar SaaS in 2026?”
what you see first
word for word
on screen, in order
There is sound here, and someone talking over it
aac, 2 channels at 44100 Hz, averaging -27.1 dBFS with peaks at -9.1.
Cut timing not measurable: there is speech here, so the beats are not reliable enough to measure cuts against
69.5k views · 23.7× this creator's median
Posted 3d ago, which is about 23.1k views a day since it went up. Their median video does 2.9k.
Band: big. Green means at least three times the median.
the account is at 2.8×
Their most recent half of videos has a median of 3.7k against 1.3k for everything before it, across 459 videos and 1426 days.
459 videos over 1426 days. Median 2.9k views. Green is three times that or better.
every video, newest first
Showing the newest 120 of 459. The figures above are computed over all of them.
files
what is actually in the mp3
33.2s · 192 kbps · 44100 Hz · ID3 v2.3 · cover from the platform's own thumbnail. These were read back off the finished file, not copied from what was asked for.
Gaurav launched Fastlane just two months ago. It’s already doing $69K...
starterstory
TikTok
Starter Story
2026
original sound
https://www.tiktok.com/@starterstory/video/7683952455289146637?_r=1&_t=ZP-99gHmjnL3mh
123 KB embedded
measured off the file
what it was probably made with
Everything below this line is reasoning on top of those numbers rather than something read off the file. It is a guess with its working shown, and it is labelled that way on purpose.
The whole clip is one shot (0 hard cuts, average shot length equals the full 33.2s duration), so this was most likely filmed as a single continuous vertical take on a phone rather than assembled from multiple clips in a multi-cut edit.
Only 6 of the 51 on-screen text regions move with the footage, meaning the text is mostly static overlay rather than content scrolling under a moving camera, which makes a screen-recording origin (e.g. a captured browser or app session) unlikely.
Speech fills 0.99 of the runtime with a longest inter-word gap of just 0.17s and a fast, even 207 words per minute, a cadence tighter and more metronomic than typical live-mic speech, which points toward a synthetic TTS voice track (tools like ElevenLabs) rather than a recorded human read.
27 of the 51 text regions echo spoken words, confirming burned-in captions generated from the audio and rendered into the frame, the standard output of an auto-caption feature like CapCut's word/phrase-pop caption presets.
Loudness range is narrow at 3.2 LU with a true peak of -9.1 dBTP, consistent with a single dry voice track and no music bed layered under it (a music mix would typically push the true peak closer to the -1 dBTP ceiling).
The file is only 0.38 MB at 92 kbps overall for a 720x1280 HEVC/AAC export, indicating aggressive compression for a small file size, i.e. an app-side re-encode rather than a high-bitrate export master.
how you would make one
- Shoot (or select) a single continuous vertical 9:16 clip on a phone camera, matching the one-shot, no-cuts structure.
- Write a short script sized to roughly the target word count and duration you want (this file ran 114 words over 33s).
- Generate the narration with a text-to-speech engine such as ElevenLabs, or record a voiceover directly if a human read is preferred.
- Import the video and the voice track into a mobile/desktop editor such as CapCut, and trim/align the audio to the footage.
- Run the editor's auto-caption/transcription tool on the voice track to generate word-level timed captions.
- Apply a word-by-word or phrase-pop caption style/preset and position it in frame, since these captions get burned into the video rather than shipped as a separate subtitle track.
- Normalize the audio mix toward the short-form platform loudness target during mixdown rather than leaving it at an arbitrary level.
- Export vertical at your target resolution and frame rate, choosing H.264 for broad compatibility or HEVC for a smaller file size.
- Play back the export at full size before publishing to confirm captions stay synced and legible against the footage.
what is easy to get wrong
- At 207 words per minute with gaps as tight as 0.17s, caption timing needs per-word timestamps, not just phrase-level timing, or captions will visibly drift or overlap the audio.
- This file measures -24.5 LUFS integrated with a -9.1 dBTP true peak, well under the roughly -14 LUFS / -1 dBTP target that TikTok, Reels, and Shorts commonly normalize to, so a re-export using this file's loudness as a reference would end up sounding quiet after the platform's own normalization pass.
- The frame is 720x1280 at 25fps, which is lower resolution and a different frame rate than the 1080x1920 at 30fps commonly recommended for vertical exports, so matching this file's specs rather than platform-recommended specs risks a softer-looking result.
- With only 6 of 51 text regions moving, don't default to screen-recording technique (capture + scroll) for a shape like this; the near-fully-static text points to overlay captions on a normal camera shot instead.
- 92 kbps overall is very low for 720x1280 video; treating this bitrate as a target during export is likely to reproduce visible compression artifacts rather than a clean master.
tools checked against the web
4 live web searches ran for this.