Why we don't do one-click summaries
Most "AI video summarizer" tools collapse two distinct steps into one and lose accountability in the process: you paste a URL, you get a paragraph, and the source transcript is hidden somewhere you cannot see it. When the summary is wrong — and large language models are still wrong about specific claims roughly some of the time — there is no obvious way to check what they invented.
We keep the steps separate on purpose. shrybe produces a clean, timestamped transcript — from a YouTube link, a podcast episode, or an audio file you already have. You decide which AI tool, which depth of summary, and which prompt. The transcript stays with you as the verification layer. That is the only workflow that scales honestly.
The two-step in practice
- Transcribe with shrybe. Paste the YouTube link, podcast URL, or audio file URL into the form above. Whisper writes a timestamped transcript at about real time on the free lane, so a 10-minute video takes about ten minutes (the paid fast lane does it in under ten seconds). Click Copy or Download .txt.
- Paste into your AI of choice. Open ChatGPT, Claude, Gemini, or any other frontier model. Paste the transcript. Ask for the summary you want — bullet list, TL;DR, 5,000-word essay, key quotes.
- Verify against the timestamps. The original transcript still has the time codes. For any specific claim in the summary you care about, scrub the source at that time. The whole verification loop takes seconds.
Prompts that actually produce useful summaries
Prompt quality matters more than model choice for transcript summarization. The pattern that works across all three frontier models:
- Tight TL;DR. "In 50 words or fewer, summarize this transcript. Preserve any specific numbers, dates, or names mentioned."
- Bullet list with timestamps. "List the 5–7 most important moments in this transcript as bullet points. For each, include the start timestamp from the transcript."
- Quotable lines. "From this transcript, extract the 3–5 most quotable sentences — the ones a journalist would highlight. Quote them verbatim with their timestamps."
- Long-form essay. "Convert this transcript into a 1,000-word essay in the speaker's voice. Preserve their argument structure. Do not invent claims they did not make."
How frontier AI models differ at this task
All three current frontier models — OpenAI ChatGPT, Anthropic Claude, Google Gemini — are good enough for transcript summarization. The differences are marginal and stylistic:
- Claude writes tighter prose, is more willing to say "I do not know," and tends to flag when a claim is the speaker's opinion versus stated fact.
- ChatGPT is the fastest and best at structured output. Tables, markdown, JSON — it produces them cleanly on first try.
- Gemini integrates with Google Drive and Docs, which matters if you push summaries into existing documents.
Pick the one whose tone you like. The summary will improve more from a better prompt than from a different model.
About hallucination — the honest part
Modern frontier models still hallucinate, and so does the transcription step in front of them — OpenAI's own model card warns that Whisper's predictions "may include texts that are not actually spoken in the audio input". We are not going to quote you a hallucination rate. The published figures shift with every model release, they are measured on benchmark tasks rather than on whatever you are about to summarise, and a number borrowed from one lab's announcement would tell you nothing about your video. What matters is the shape of the risk: a long summary will probably contain at least one claim the source does not support, and you will not catch it by reading the summary again. The transcript-then-summarize workflow exists exactly because you can catch those one-or-two wrong claims by skimming the source. A one-click summarizer offers nothing to skim against.
Frequently asked
Why not a single one-click summary tool?
Two reasons. One: the transcript is the trust layer. Single-click summarizers hide the source text, so when the summary is wrong — and they are, sometimes — you have no way to verify. Two: different summary needs ask for different prompts. A 50-word TL;DR, a structured outline, and a list of quotable lines are three different jobs. Splitting the steps lets you do whichever job the moment requires, without us guessing.
Which AI tool gives the best summaries?
Honestly, all three current frontier models (ChatGPT, Claude, Gemini) handle transcript summarization well. Claude tends to write tighter prose and is more honest about uncertainty. ChatGPT is faster and better at structured output (markdown lists, tables). Gemini integrates with Google Drive if you want to push the summary to a doc. Try one. If you do this regularly, pick the one whose tone you prefer.
Can shrybe summarize the transcript itself in the future?
Probably, on the paid tier. We have not shipped it because running an LLM call on every transcription is expensive, and forcing one model's voice on every summary is the wrong trade-off when ChatGPT or Claude in a separate tab does the job under the user's full control. If we ship summarization, it will be optional, configurable, and honest about which model produced it.
What length of summary should I ask for?
Depends on the source. A 60-minute interview generally compresses well into 300-500 words without losing structure. A short 5-minute talk does not need a 500-word summary — ask for 100 words or a bullet list. A good prompt: 'Summarize this transcript in roughly 5 bullet points, preserving any specific numbers or quotes.' Specific instructions produce better output than 'summarize this'.
How accurate are AI summaries — do they hallucinate?
Modern frontier models hallucinate less than they used to, but they still do. The mitigation is the transcript itself: it stays with the timestamps, so you can spot-check any specific claim. The most common failure mode is confident over-summarization — the model invents a unifying theme that the speaker never explicitly stated. Reading the first and last paragraph of the source transcript usually catches this in under a minute.
Is this workflow better than using a "YouTube summarizer" Chrome extension?
It is honest. Most Chrome-extension summarizers either rely on YouTube's own auto-captions (which are inaccurate on accented speech, jargon, and proper nouns) or run a tiny model locally that drops detail. The shrybe-then-LLM workflow gives you a higher-quality transcript and lets you choose the summarization model. The trade-off is that it is two steps instead of one. For occasional use, the extension is convenient; for content where accuracy matters, the two-step is better.
Will the summary respect copyright?
Summarization for personal note-taking, study, or research generally falls under fair use in most jurisdictions. Republishing the summary as your own content, or distributing it commercially, is a separate question — that is between you and the original creator. shrybe is the transcript layer; what you do downstream is on you.
Start with the transcript.
Paste your video or podcast URL above. The transcript is your starting point; from there, the summary is whichever shape works best.