The video-production plugin started life as one video project tuned by hand, and when I turned that work into reusable scripts I checked them against the original byte for byte. The cut lists, the face and camera tracks, and one finished short’s audio, captions and decoded picture all matched exactly.
What it does now is take raw talking-head recordings from a phone and turn them into a themed series of captioned short videos plus a longer cut, running on the person’s own Mac. I made it for people with plenty to say to a phone camera and no interest in learning to edit.
Sound and words come first
The first stage measures the audio and looks for wind in the pauses between words, where it shows up as low rumble that keeps going when nobody is speaking. It cleans the voice with a noise-removal model and keeps two versions, an aggressive one used only for transcription and a gentler one for the edit, so the pauses in the finished video don’t sound dead. Then it transcribes the speech and force-aligns every word to the waveform. On real footage the transcriber’s own timings put 13 to 15 percent of words starting outside detected speech, and alignment brings that under 1 percent, which is what keeps captions and cuts landing on the right syllable.
After that, doubtful words get settled with evidence. Any line that says something about a person or an organisation is treated as doubtful by default, however confident the transcript looks, and the skill scores competing readings of that span against the audio, including readings that flip the meaning, like “a responsible” against “irresponsible” or “can” against “can’t”. If a meaning-flipped version fits the audio as well as the original, the run stops with a meaning check, and the line stays out of every edit unless the speaker settles it. I added those automatic flips after a test run, a fresh Claude session given only these plugins and two short real recordings, quoted a line about a person as the opposite of what the audio supported, because nobody had scored the opposite reading. A word the skill can’t resolve never goes on screen.
A story the speaker agrees to
The story-edit skill reads the verified transcript and finds the themes and the thread running through them on its own, then proposes them with quoted evidence before anything is cut. The speaker refines the themes, and only then does it order the series, pick hooks that work on muted autoplay and write the cut specs. Before any render, two independent reviewers challenge the plan: an editor-strategist and a risk reviewer who looks for how clients, investors or employers could misread a video, for splices that shift meaning and for privacy problems. Every splice has to keep the meaning the words had where they were said, tone included, so a joke can’t land in a serious context and a “sometimes” can’t get trimmed off a claim. The whole plan goes on a page the speaker can read in full before anyone asks them to approve it.
Finishing, then checking every file
Rendering smooths out handheld shake, frames each shot around the face and mouth, adds bold animated word-highlight captions and a light grade, finishes the sound with loudness normalisation, and exports 9:16 and 4:5 versions. On the handheld selfie shorts measured so far, stabilisation cut high-frequency shake by 28 to 49 percent per video at the median, which leaves them smoother and still short of tripod-steady.
The QA skill runs before anyone sees a video. It re-transcribes each render and compares it with the words the plan intended, confirms the captions never cover the speaker’s mouth, checks sync, loudness and true peak, and makes sure location metadata has been stripped from the file. Transcription, cleanup, tracking and rendering all run on the Mac, so the recordings are never uploaded anywhere for editing and nothing costs money per use.
Limits, written down
The first version went in on September 13, 2026, and the current one is 1.3.2. Setup is two messages pasted into the Code tab of the Claude desktop app, with a toolkit of about 3.2 GB that downloads once. The limits are written plainly in the README: it runs on Macs only and has been tested on Intel so far, it accepts vertical SDR recordings, it exports 9:16 and 4:5 with no 16:9 yet, it’s tuned for English, and it has one caption style with no music bed. If you have a folder of phone recordings you never got around to editing, that’s exactly the material it was built for.