← sources

Built a Claude Code skill that measures video frame by frame: same accuracy as /watch at about half the cost (216-run benchmark, MIT)

source: reddit (reddit:ClaudeCode)published 2026-10-04 10:19ingested 2026-10-05 02:40

retained text

**What it is.** video-lens is a free, MIT-licensed Claude Code skill for macOS. Claude can't watch video, so video skills hand it frames. video-lens measures every frame first (UI motion timing and easing as CSS, scene cuts, on-screen text, speech with timestamps) and gives Claude numbers and text, with images only where a look is needed. Nothing is uploaded: it runs on ffmpeg, OpenCV, macOS Vision and on-device speech recognition.

**How Claude helped.** I built it with Claude Code on Opus 5.5: the spec, the Python analysis code and Swift helpers, 60 self-tests on synthetic clips with known answers, the benchmark harness and the scorer. My part was deciding what to measure and checking the results against answer keys.

**Benchmark.** 9 tasks scored against answer keys (CSS animations rendered in headless Chrome, Korean lectures, a real screen recording), 27 runs per condition, Claude Opus 5.5, one run at a time:

| Condition | Mean score | Worst run | Cost per run | Time per run |
|---|---|---|---|---|
| Opus alone | 0.949 | 0.167 | $0.86 | 420 s |
| With video-lens | 0.956 | 0.800 | $0.66 | 265 s |
| With /watch 0.3.2 | 0.956 | 0.467 | $1.20 | 505 s |
| With video-use | 0.969 | 0.467 | $1.57 | 487 s |

**What I learned**
- On a strong model the skill barely moves accuracy (0.949 to 0.956). What it changes is cost, time and how bad the worst run gets. On a weaker model the gap is bigger: Grok 4.7 went from 0.819 to 0.942.
- With /watch and video-use, Opus usually also wrote its own ffmpeg or OpenCV analysis on top of the skill (23 and 27 of 27 runs, against 2 of 27 with video-lens). That is where most of their extra cost and time went.
- I had Opus rebuild two viral launch-week showreels as HTML from the video alone. Without the skill the rebuilds looked slightly closer to the originals (SSIM 0.82 and 0.84 against 0.79 and 0.81); with it they cost 24% less. The side-by-side videos are on the site.
- macOS only, Python 3.10 or newer.

**Try it** (free)

    /plugin marketplace add Junhan2/video-lens
    /plugin install video-lens@video-lens

Repo: https://github.com/Junhan2/video-lens
Results, raw data and the scripts to rerun them: https://junhan2.github.io/video-lens/

Feedback on where it breaks is very welcome.

document fetched 2026-10-05 00:35 · canonical https://reddit.com/r/ClaudeCode/comments/1wxc9ue/built_a_claude_code_skill_that_measures_video