Transcript-first editing fits speech better than visual storytelling
video-use gives a shell-capable coding agent a video-editing skill plus helper scripts. The agent inventories raw takes, proposes an approach, waits for approval, and writes an edit under <videos_dir>/edit/. Its main reading surface is a packed transcript with word timestamps, speaker labels, and audio events. Visual timeline composites are generated only when a cut or retake needs inspection.
That representation is well matched to talking heads, interviews, tutorials, and launch clips where spoken content determines the structure. It is less convincing for dance, sports, cinematic montage, or product footage whose meaning lives in motion and framing. The README says the model never watches the video continuously. It reasons from text and requests filmstrip-plus-waveform images at selected moments. Buyers should treat that constraint as a product choice, not magic vision.
The 508 MB install excludes external editing services
Our commit 92c2b34 checkout was small: 33 files, about 1,946 source lines, and 0.7 MB. Installing 58 Python packages succeeded in 28 seconds and occupied 508 MB. The build completed in 8 seconds. That is approachable for an agent skill, although the installed environment is much larger than the source because media tooling brings substantial dependencies.
A usable workstation also needs FFmpeg. The default transcription path requires an ElevenLabs API key and sends each source through Scribe once for word-level timestamps, diarization, and audio events. yt-dlp is optional for online inputs. Animated overlays can introduce HyperFrames, Remotion, Manim, or PIL, so an edit using those features has a wider runtime than our base package measurement.
What happened when we ran it
Our sandbox installed the 58 Python packages in 28 seconds on 3 CPUs with 8 GB of RAM, then built the project in 8 seconds. Pip-audit found 0 known vulnerabilities. No tests script or target was present, so the lab skipped tests rather than pretending a build was a test run.
The repository scan found no CI workflow files, no Dockerfile, and no tests directory at commit 92c2b34. We did not provide footage, an ElevenLabs credential, or an editing prompt, and we did not render a video. The measurements therefore cover dependency setup and packaging only. They do not establish transcription accuracy, cut quality, rendering speed, subtitle correctness, or compatibility with a particular codec.
Strategy approval is the most important control
The documented session asks the agent to inspect sources, propose an editing strategy, and wait for the user's OK before execution. That pause matters because a transcript can reveal filler words and false starts without knowing which hesitation carries personality or which imperfect take is emotionally better. A useful strategy should name the chosen takes, intended length, pacing, subtitle treatment, and any generated overlays before FFmpeg touches the cut.
The pipeline then transcribes, packs, reasons, writes an edit decision list, renders, and evaluates cut boundaries. A self-check can rerender up to 3 times when timeline images expose visual jumps, audio pops, or hidden subtitles. This catches mechanical faults that the helpers know how to inspect. It cannot decide whether a joke lands, a brand claim is approved, or a speaker looks awkward in the chosen frame.
FFmpeg portability is still active work
Recent pull requests address preserving source frame rate, rotation metadata in portrait detection, apostrophes in concat lists, Windows path and UTF-8 handling, selecting the right audio track, and subtitle burn-in when Homebrew FFmpeg lacks libass. Issue 125 reports a Unicode encoding crash on non-UTF-8 Windows output. Issue 121 asks for a doctor command to verify FFmpeg and optional overlay dependencies.
Those changes describe the edges of programmatic editing: filenames, fonts, codecs, metadata, console encodings, and FFmpeg builds differ across machines. Run a preflight on the actual editing host. Include portrait and horizontal clips, rotated phone footage, apostrophes and non-ASCII characters in paths, silent clips, multiple audio tracks, and the subtitle font you plan to ship.
The skill belongs inside an agent permission model
Manual setup symlinks the repository into ~/.claude/skills/video-use for Claude Code or ~/.codex/skills/video-use for Codex. The setup prompt also supports Hermes, OpenClaw, and other agents with shell access. That makes video-use a Claude Code tool and requires the claude-code tag, but it is not a standalone desktop editor with its own permission dialog.
The agent can read footage, create files, run FFmpeg, call transcription services, and install optional packages. Scope it to the project folder, keep the API key in an environment file outside conversation, and review commands that download or execute third-party assets. Preserve raw footage and treat edit/ as disposable output until a human approves the final file.
August activity is high, while releases and tests are absent
GitHub listed 21,387 stars and 79 combined open issues and pull requests when fetched. The repository was pushed August 26, 2026, and active work that week covered a native macOS EDL editor, frame-rate handling, portrait rotation, Windows encoding, audio selection, and subtitles. The latest-release API returned no release.
That pace makes video-use interesting and unsettled. Its 8-second build lowers the cost of trying it, while the absent test target means each workstation and media set carries more validation work. Use it on copies of raw footage, start with a short speech-led clip, and compare the proposed edit with the rendered result. If the transcript is not the natural outline of your video, a conventional timeline will be faster to trust.