How Claras works

Claras does one thing first and everything else after: it gets the transcript of the video. The summary, the chapters, the Ask AI answers and the exports are all built from that text. Here is each step in order, as it happens in the web app and the Chrome extension.

1. You add a video

You sign in, then paste a YouTube, Instagram or TikTok link into the web app, or upload a video or audio file. In the Chrome extension, you click Summarize under a YouTube video. You also choose the language you want to read in, one of 15.

Before any work starts, Claras checks your plan. That check doesn't use up anything. A video only counts once its analysis is finished and saved, so a run that fails along the way costs you nothing.

2. Claras reads the captions

Most YouTube videos come with captions, written by the creator or made automatically by YouTube. When they exist, Claras fetches them with their timestamps. This is the quick path, with no AI involved yet.

Which captions it takes depends on where you are. The web app asks for English first. If the video has none, it tries English from another region, such as British English, then one of the other caption tracks the video has. In the extension, you can pick the caption language yourself from the menu next to the link.

Then Claras checks the text before trusting it. It looks at the alphabet and at common everyday words to make sure the captions really are in one of the video's own languages. A machine translation into some other language is turned down, and the next option is tried. A copy that passes is kept, so the next person who opens the same video gets it faster.

The transcript is never translated or corrected. If YouTube's automatic captions mishear a name, the transcript has the same mistake, and so does anything built from it.

3. No captions? AI listens to the audio

Some videos have no captions at all, and Instagram, TikTok and your own files never come with any. For those, Claras has the audio transcribed by AI. How that works depends on the source.

YouTube videos without captions

Claras first checks how long the video is. On the free plan, AI transcription covers videos up to 30 minutes, and only when the length can be checked. On a paid plan it covers videos up to 3 hours. A long video can take longer than one request is allowed to run. When that happens, you see a message asking you to try again in a minute or two. Claras keeps track of the job, so your retry picks up the same transcription instead of starting over.

Instagram and TikTok

Claras looks up whether the video is public, whether the post is a video at all, and how long it runs. Links up to 30 minutes work, and the length has to fit the transcription hours left on your plan. Then the audio is transcribed and the words come back with their times.

Your own files

Here the work starts in your browser. It reads the file, drops the picture, and cuts the sound into slices of up to 10 minutes. Each slice overlaps the next by a few seconds, so no sentence is lost at a cut. The slices go off four at a time to be transcribed. The first slice sets the language for the rest. Claras then joins the slices on the file's own clock, so the times match your recording.

The file itself never leaves your computer and isn't stored. Only the audio slices are sent, and they are dropped once transcribed. Finished slices are kept in your browser, so if the tab closes halfway, picking the same file again carries on from where it stopped.

4. The video is read in parts

A long transcript is too much to read well in one go. Ask an AI to summarize three hours at once and the last two hours tend to get a line each. So Claras cuts the transcript into stretches of 20 minutes and reads each one in full.

Up to four stretches are read at the same time. Each one is sent with the last minute of the stretch before it, so a thought that crosses the line isn't cut in half, and with the opening lines of the video, so the AI knows what the whole thing is about. Each stretch comes back with two things: a summary of that stretch, and chapters for it.

The chapters are then checked. Every chapter time is moved to the start of the nearest caption line, so each chapter begins on a line that was really spoken. A chapter that doesn't hold its topic for long enough folds into the one before it. That minimum grows with the video, from 45 seconds for anything up to an hour and a half to two and a half minutes for five hours or more.

Last, Claras looks for holes. If more than five minutes of the video ended up with no chapter, it reads that stretch again, up to the three biggest gaps.

5. The summary is pulled together

A final pass reads the chapter list and the section summaries, not the whole transcript again. That keeps it much smaller than the transcript, even for a five-hour video. From those, it writes the opening and the Key Takeaways, suggests three questions you might ask, checks the title against what was actually said, and marks chapters that look like a sponsor read or an ad.

All of this is written by GPT-4o Mini, the model Claras runs on every plan. Every request tells the AI to answer in the language you picked, whatever language the video is in. That is how a Japanese summary of an English lecture comes about while the transcript stays in English. On a paid plan, the Chrome extension can send these requests to your own OpenAI, Anthropic or Gemini account instead.

If you added your own instructions in Settings, such as "Use short bullets", they are added to the rules for the section summaries. They change the style, not the steps.

6. Ask AI finds the right part

When you ask a question, Claras doesn't send the whole transcript to the AI. It splits the transcript into short sections, each roughly a minute or two of speech and marked with its start time. It picks the five sections that share the most words with your question, and sends those with the video title and your last few messages, so follow-up questions make sense.

The answer usually gives a time, like [02:15], where the point comes from, so you can check it yourself. Because sections are matched by the words they share with your question, using the speaker's own words gets better answers. For questions about the whole video, the summary is the better place to start, since it was built from every part.

After a Bulk Summaries run, the same idea covers the whole batch: each answer reads ten sections across all the videos, and the AI is asked to say which video each point came from.

7. Many videos at once

Bulk Summaries starts from a public playlist or a YouTube search page. Claras lists the videos, you pick the ones you want, and it fetches each transcript the same way as above. A combined summary then reads the start of each transcript within one shared budget, so the more videos you pick, the less of each it reads. A summary per video reads much more of each one but doesn't compare them.

8. What is saved

When the analysis is done, the transcript, the summary and the chapters are saved to History in your account. The web app and the extension share it. Reopening a video from History is free and runs nothing again. Ask AI conversations aren't saved. For uploads, the result is saved but the file isn't, so to play it again you choose the file again.

What this means in practice

• Claras knows what was said, not what was shown. Slides, code, charts and on-screen text don't reach the transcript, so a mostly silent video gets a thin result.
• Good captions, good results. Caption mistakes carry through to the summary and the answers. Check a name against the transcript before you quote it.
• No speaker names. Captions rarely say who is talking, so in a panel a point can be pinned on the wrong person.
• Long videos get the same attention at the end as at the start, because every 20-minute stretch is read in full.

Each step has its own page with more detail: transcripts, AI summaries, AI chapters, Ask AI and languages. If something goes wrong along the way, the help center lists the messages you might see and what to do. To try it on a video of your own, create a free account.