AI Tools: Claude Can Now Watch Videos (100% Free)

Watch on YouTube: https://www.youtube.com/watch?v=mOy7JQjTYwE

Channel: Moritz | AI Systems

Introduction

The video from Moritz | AI Systems shows how to give Claude the ability to “watch” YouTube videos without relying on its native text‑only processing. By installing an open‑source skill called eyes, Claude can index a video’s frames, transcript, on‑screen text, timestamps, and audio/color data. Once indexed, the model can answer detailed questions about the video quickly and consistently, at a cost of only a few cents per video.

What the Video Covers

The presentation walks through the limitations of Claude’s built‑in capabilities, explains how the eyes skill works, demonstrates its core features, provides cost estimates, and outlines practical use cases. Key points include:

• Claude cannot natively process video; it only handles text and static images.
• The eyes skill extracts frames, transcript, on‑screen text, timestamps, and audio/color data to create a local index.
• Indexed data enables fast, repeatable queries for search, burst analysis, layout detection, and video DNA comparison.
• Costs range from approximately 1 cent for a short clip using a cheap model to about 30 cents for a 20‑minute video processed with the top model.
• Installation is performed via a GitHub repository; after setup, the agent invokes the skill with the command /eyes <YouTube‑URL>.
• Use cases highlighted are tutorial step extraction, competitor video analysis, clipping, style copying, and edit validation.
• The tool improves consistency compared with Claude’s ad‑hoc FFMPEG approach, which can vary depending on the prompt.

Step‑by‑Step Walkthrough

1. Install the eyes skill – Copy the GitHub repository link from the video description and instruct your agent to install it. This step sets up the necessary dependencies and makes the /eyes command available.
2. Index a video – Run /eyes <YouTube‑URL> pointing to the target YouTube video. The skill will:
   • Extract frames at a configurable rate (the demo mentions bursts of 8 frames per second).
   • Download the transcript and any on‑screen text.
   • Capture audio and color information for later analysis.
   • Store all extracted data locally so subsequent queries are fast and consistent.
3. Run queries – Once indexing is complete, you can ask Claude specific questions. Examples from the video include:
   • Searching for on‑screen text or keywords.
   • Using the burst feature to capture a rapid series of screenshots from an interesting segment.
   • Applying the measure command to generate an editing DNA report (e.g., cuts per minute, percentage of shots under one second).
   • Comparing two videos by analyzing their DNA reports to spot similarities or differences in editing style.
   • Cropping or extracting specific frames for further use.
4. Apply insights – The DNA report can be used to:
   • Edit a raw clip to match the pacing and style of a reference video.
   • Validate your own edits before publishing by checking against reference metrics such as layout changes or shot length distribution.

Tools, Numbers, and Technical Details

The video provides concrete figures that help estimate the expense of using the eyes skill:

• ≈ 1 cent for a 45‑second video processed with a cheap model.
• ≈ 2 cents for the same length with a mid‑tier model.
• ≈ 10 cents for a 45‑second video using the top model.
• ≈ 20 cents for a 2‑minute video with the top model.
• ≈ 30 cents for a 20‑minute video with the top model.

Other technical mentions include:

• The skill relies on FFMPEG for frame extraction when used ad‑hoc, but the eyes tool provides a consistent, repeatable alternative.
• Burst mode captures approximately 8 frames per second.
• The measure feature can report metrics such as 70 cuts per minute and T7% of shots lasting under one second.
• Installation source is the public GitHub repository linked in the video description.

Caveats and Hype Check

“Claude can’t actually watch videos. It will kind of rebuild everything from scratch.”

The presenter emphasizes that the tool is free and open‑source, with no affiliate promotions or hidden revenue claims. Cost estimates are approximate and depend on the model selected; the core functionality is presented as genuinely accessible without upsells. However, users should note that:

• Initial indexing requires processing time proportional to video length and the chosen model’s speed.
• Local storage is needed to hold the indexed frames, transcript, and audio data.
• While the skill improves consistency over ad‑hoc FFMPEG prompts, the quality of on‑screen text extraction depends on the underlying OCR capabilities.

Concrete Next Actions

To start using the eyes skill with Claude, consider the following steps:

• Obtain the GitHub repository URL from the video description and run the installation command via your agent.
• Test the skill on a short YouTube clip (e.g., a 45‑second tutorial) using /eyes <URL> to verify indexing and query response times.
• Experiment with the burst feature on a segment containing fast motion or animation to see how many frames are captured per second.
• Generate a measure report for two competing videos and compare their editing DNA (cuts per minute, shot length distribution) to identify stylistic differences.
• Search your indexed library for specific on‑screen text or keywords relevant to a research task.
• Apply the DNA report to edit a raw clip, aiming to match the reference video’s pacing and layout changes.
• Before publishing any edited video, run a validation check against the reference metrics to ensure consistency.