Tool — ViralClipAI
Give it a transcript. It scans for hook moments, emotional beats, and tactical takeaways, then returns the three segments most likely to hold a viewer — each with a start and end timestamp, a viral score with a reason, a hook overlay for the first three seconds, and a caption with hashtags. Nothing is invented: every timestamp and every word comes from your transcript.
01 — The rules it runs on
The agent prompt is deliberately short. Most of it is spent saying what it is not allowed to do, which is what keeps the output trustworthy enough to publish.
Short-form video strategist for TikTok, Shorts, and Reels. It knows the shape of a clip before it reads a word of the transcript.
Find high-hook moments, emotionally resonant statements, tactical takeaways, and surprising insights. Pick the top three. Score, timestamp, caption, and overlay each one.
Fixed structure, every time: clip title, start, end, duration, viral score, visual hook text, social caption. Same shape on every run means the output is diffable.
A hard 15–60 second band. No fabricated dialogue. No shifted timestamps — every one lands on a real cue boundary in the source transcript.
Empty, corrupted, or transcript-free input gets one fixed sentence and no output. It never improvises to fill a gap.
02 — Real output
Run against a real 25-minute video — a marketing strategist building a plan out loud. Nothing below is a mock-up. The three clips come from three different parts of the video, spaced so they don't repeat the same passage.
Source Watch Me Build a Marketing Strategy in 20 Minutes For a Completely Random Business 25:15
Why it scored 62
It sets up “here's what” and holds 32s without a dead stretch, so the payoff lands before the scroll.
Visual hook text first 3 seconds
Social caption
So, here's what we're going to do next.
Signals detected
Why it scored 67
A 32s window that runs one thought start to finish — modest hook density, but nothing to trim.
Visual hook text first 3 seconds
Social caption
I mean, these things are terrifying.
Signals detected
Why it scored 77
A 32s window that runs one thought start to finish — modest hook density, but nothing to trim.
Visual hook text first 3 seconds
Social caption
So, we're not talking about massive, massive amounts of money, which means that it's very likely that their success of their
Signals detected
Every timestamp above is a real cue boundary from the source transcript, and every quoted phrase is a real verbatim excerpt. The code is on github.com/arlinda-k14/.hubsy.
03 — How it runs
Takes a YouTube URL or a timestamped .txt, .srt, or
.vtt transcript. If there is no usable dialogue, it refuses with the exact
sentence and stops.
Every 15–60 second window is scored on hook, emotion, tactical, and surprise density, then filtered for sentence boundaries, filler words, and speech rate. Peak retention sits near 32 seconds.
The top three non-overlapping windows are trimmed with ffmpeg into
clips/, optionally cropped to 9:16 for Shorts, Reels, and TikTok. Each render
is probed afterwards to confirm the duration matches the blueprint.
04 — Limits
A tool that only lists strengths isn't telling you anything useful.
05 — Contact
Same pattern, different pipeline — repurposing a webinar, clipping customer interviews, turning a podcast into a month of posts. Tell me what's not working and I'll tell you what it would take.