How it works
How Garfunkel analyzes a video
Paste a link and Garfunkel turns the video into a structured report: what’s said, what’s shown, the hook, the same set of tags as every other post and, if you ask, how people reacted. Here is each step, in order.
Copies this page as Markdown, ready to paste into an AI assistant.
Step 01 of 08
Getting the post
We open the link and fetch the video together with what the platform shows about it: the caption, the creator, and the views, likes, comments and shares it reports. Those numbers are kept exactly as reported, so you can sort and compare posts by them later. The video file is only kept while we work on it and is deleted once the analysis is finished.
Step 02 of 08
Splitting the video
We check whether the video has a soundtrack, then separate the sound from the picture so that each can be examined on its own. The sound goes on to be transcribed and the picture goes on to frame sampling, and the two come back together when the post is summarized.
Step 03 of 08
Transcribing the audio
A speech model writes down everything said in the video, with a timestamp on every line, so you can see what was said and when. A video with speech is never analyzed without its transcript: if this step fails, it is retried rather than skipped, and a post that fails is never charged.
Model: Qwen3-ASR-0.6B · an open-source speech recognition model
Step 04 of 08
Choosing the frames
We use a frame sampling algorithm to reduce the number of frames while keeping the overall meaning of the video intact. The frames that remain are what the AI model looks at, together with the transcript, in the steps that follow.
Step 05 of 08
Summarizing
An AI model reads the transcript and the caption while looking at the frames, so it understands the post as a whole rather than one piece at a time. It writes a short summary and the story in order, beat by beat, with each beat tied to a moment in the video.
Model: gpt-6.1-sol · reads the transcript and the sampled frames together
Step 06 of 08
Finding the hook
The model looks closely at the first few seconds to name the hook: the words or image that grab attention, what kind of hook it is and the moment it lands. It also notes the promise the hook makes to the viewer, how the video pays it off and why it works.
Model: gpt-6.1-sol · the same model and the same pass as the summary
Step 08 of 08
Reading the comments
When you ask for comments, we collect them and the model reads each one in the context of the post: its caption, summary and transcript. That way a reply is judged against what the video actually said and showed, and the report shows who agrees, who pushes back, what people ask and what they liked.
Model: gpt-6-luna · reads each comment against the post
How the context stacks
Each step builds on the one before it, so the later steps never work from the raw video alone. This is what each one reads.
- 01TranscriptReadsThe soundtrack
- 02Summary and hookReadsTranscript, caption and sampled frames
- 03TagsReadsSummary, transcript, caption and sampled frames
- 04Comment reportReadsComments, caption, summary and transcript
Every post, the same shape
Every post is analyzed into structured output with a fixed schema: the same fields, in the same format, for every post. That keeps results consistent and comparable, easy to sort and filter, and clean to export to a spreadsheet, the API or MCP.
{
"format": "Tutorial",
"theme": "Food & drink",
"emotion": ["Joy"],
"tone": ["Playful"],
"hook": { "type": "Demonstration", "start": 0.0, "end": 2.0, "text": "So here’s the trick…" },
"tags": ["home kitchen", "pan", "jump cuts"],
"transcript": [{ "start": 0.0, "end": 2.4, "text": "So here’s the trick…" }],
"comments": { "agree": 52, "push_back": 14, "questions": 18 }
}The field names here are illustrative; the docs list the real ones.
What you get
- A summary and the story in order
- A timestamped transcript
- The hook: what it is, when it lands and why it works
- The same set of tags as every other post
- A comment report, when you ask for one
- Search by meaning: find posts by what they’re about, not only by the words in them
- Exports to spreadsheets, the API and MCP
The video file is deleted once the analysis is finished.