In the first part we described how the IT Reelsmaker plugin for Claude Code edits vertical videos from a “talking head”: it cuts the recording by sound, adds subtitles and builds graphics in your brand colours. Back then the plugin was at version 1.20.0. Since then versions 1.21 to 1.38 have come out: on 13 November 2026 the core reached 1.38.1. The main news from these three weeks is that the plugin can now edit videos for Reels, Shorts and TikTok without any filming, when you’re not in the frame at all. You record the text as a voice note on your phone and get an explainer made of scenes drawn in code, and next to the scenes there can be a brand mascot that moves its mouth to your recording. The same voice note can also go on top of your own shots filmed without sound, or on top of photos. And if there’s no voice at all, you get a video with no speech: text on screen, sounds and music built in code.
The core’s principle hasn’t changed since then. The plugin generates nothing with a neural network: no images, no video, no voice. The scenes and the mascot are motion design that Claude writes as code in Remotion. The voice is always yours, recorded, and short sound effects and music are now built in code too. A voiceover in a clone of your voice exists only in the online add-on; more on it below.
Reels without filming: a video made of scenes timed to your voice
Since version 1.26.0 you may not need any video for a clip at all. A voice recording on your phone is enough; m4a, wav, mp3, ogg and opus all work. The plugin cuts that recording the same way it cut the video in the first part: it removes pauses, finds retakes, speeds the speech up slightly, levels the volume and checks the edges of the cuts. If the file is empty or too short, the plugin says so straight away instead of failing halfway through. It also warns you about a steady background under the voice, such as music or a noisy room: that background would then play under every word of the video, so it’s better to record the voice clean and add music during the edit. A repeat run with no changes is taken from the cache.
The picture is made of designed scenes (a hook, a number, a list, a diagram), and they land on the spoken words. This kind of video has a new voice setting: on-camera speech (the default) or voiceover. In voiceover mode, the subtitles stay on top of the full-screen scenes, and the scene text sits above them without overlapping. Your recording sets the length of the video. For recordings longer than about 70 seconds, the plugin warns that on a laptop with 8 GB of memory the render may not go through. Mastering will refuse to build a video whose render has lost the voice: without that check it would silently hand over a mute promo.
Who needs this? An expert who doesn’t like the camera. A series where a face adds nothing. Someone who has the idea now but won’t be able to film themselves until the day after tomorrow.
A brand mascot lip-synced to your recording
Version 1.27.0 added a mascot. It’s optional and never goes into a video on its own. Claude draws the character as a Remotion component right in your project, in the brand colours, and picks its personality from the brand’s tone and audience. Every brand ends up with a figure of its own.
The mascot stands in the bottom left corner above the subtitles. While you talk it never freezes, and on the words you mark in the plan it reacts with a gesture that fits the line. A gesture is tied to a word, like the inserts in the first part, so after a re-cut it stays on its line.
The mouth opens according to the loudness of your recording. The plugin measures the sound in every frame and counts as speech anything 10 dB louder than the background between words. In pauses the mouth is closed. Since version 1.32.0 the mouth also takes the shape of the vowel into account: it can stretch wider or get rounder. This is also measured from the sound of the recording itself, with no transcript, so a high or a low voice makes no difference. The measurement tells vowels in real speech apart only roughly, so most of the time the mouth stays neutral, but ordinary speech no longer looks “mostly round”.
Since version 1.31.0 the mascot blinks like a living character. The eyelid closes through “half shut” instead of snapping, about one blink in five is a double blink, and the pace is human: once every 2.6–4.6 seconds. The same video blinks the same way at every render. A pose change no longer looks like a jump cut: the mascot blinks on it, and a character from a ready-made image changes pose smoothly, over two frames, while its eyes are shut.
Since version 1.34.0 a drawn mascot’s gesture looks like acting rather than a hinge moving. Its parts catch up with the new pose in turn: the eyes first, the head a frame later, the arms another frame after that, then the hands. At loud moments the pose clicks into place in two frames and holds, and if the brand tone allows it, the mascot sways slightly after the click, a sway that fades out in half a second. In a calm tone the transition stays smooth, as before.
How noticeable the character is depends on the brand’s tone. With the premium tone there’s no mascot at all; with calm tones such as expert and tech it’s restrained; with the rest it works at full strength. It only “bounces” on a change of pose in a loud scene, and only for a brand that allows it. The checks keep the mascot off the subtitles, the platform’s buttons, scenes, callouts and memes. A gesture comes no more often than once every 0.8 seconds and never at the moment a scene appears, and there’s only one mascot per video. If the video has been re-cut but the mouth is still calculated from the old version, the plugin will refuse to build the project and ask you to recalculate.
A mascot from a ready-made character image
If the brand already has a character drawn by an illustrator, there’s no need to redraw it in code. Since version 1.28.0 the plugin accepts a pose sheet: a grid in one image or a folder of PNGs, one file per pose, with separate “talking” and “blinked” frames if you have them. A script cuts this into an atlas, puts it in the brand folder and builds a contact sheet.
The plugin handles the background carefully. A transparent image stays as it is. A flat background is removed from the edges inwards, so white eyes on a white background survive. Pieces of background trapped inside the figure, say between an arm and the body, are numbered on the contact sheet and removed only on your command. The plugin won’t take an image on a gradient or photo background. All poses are scaled and shifted so that the head stays in one spot and doesn’t jump when the pose changes. Since version 1.29.2 the plugin suggests where the character’s head and eyes are on its own, so you have to correct them by hand less often.
The mouth is either a drawn “talking” frame or an oval over the image; blinking is either a “blinked” frame or an eyelid in the skin colour. An oval over a finished drawing is noticeably worse than a drawn frame, so if you’re commissioning a character from an illustrator, ask for that frame straight away. An asymmetrical character, with a logo on its T-shirt or hair parted on one side, is never mirrored by the plugin.
Callouts on objects now follow the object
In the first part, callouts (a dot on an object, a line and a label with its name) only worked in a static shot. Version 1.21.0 removed that limit. On a pan or a handheld walk-through of a flat, the plugin tracks the object frame by frame and moves the dot with it. At a cut and where the object leaves the frame, tracking stops and the dot fades out smoothly. For a slow pan you can set anchor points by hand. The plan check measures the picture along the whole path and won’t pass a callout that has slid off its object, and after a re-cut it will ask you to recalculate the tracking. This needs OpenCV and numpy.
There are more kinds of callout now: a caption right above the line with no label box (where the background is too light for such a caption, the plugin won’t place it), a badge with a price or rating in a circle, a ring around one object among others, and a dimension line between two points with the size in the middle. Any number on a callout still has to have a source. A callout can also go on a person: the dot sits on the shoulder, follows it, and the caption moves to the free side of the frame. All of this is still experimental.
An AI agent checks the render before you see it
It used to be that a person saw the finished render first. Since version 1.23.0 the agent watches it first. It looks at every cut close up and at the whole video at phone-screen size, and writes its notes as “time · where · problem → fix” lines with priorities. After two rounds of fixes it doesn’t silently fix what’s left a third time; it hands the video over as “passed with notes”.
Nothing is swapped out silently. If a program, a model or a font is missing, or some step fails, the agent says what’s missing and what that will change in the video, and only then takes a simpler route.
Separate work went into fighting the template look. The plugin finds stretches longer than 3.5 seconds where nothing changes in the picture while it’s still at the plan stage. A promo made only of scenes is checked for being a slide show: one scene type everywhere, three identical ones in a row, no main moment. A “swap test” catches a video that would fit any topic. And the rules of a new brand come with a short list of the things that make a video look like everyone else’s.
Sound effects in code, and music from the track’s busiest part
In the first part we offered a starter pack of 260 CC0 sounds. Since version 1.24.0 you don’t need it: the plugin synthesises short sounds for on-screen events itself through ffmpeg, using recipes from the skill. There’s nothing to download and no licence to check. Only events you can see in the frame get a sound. It lands one frame after the picture, and a repeated sound alternates between two or three variants, so you don’t hear the exact same sound on every cut.
Music no longer starts with a quiet intro. If the video has no single moment the track must be fitted to, the agent looks for the video-length stretch of the track with the most energy and starts from there. A series of videos now has a decision log: a choice made on one episode carries over to the next ones, even if each is edited in a new conversation.
Picture separate from sound: the groundwork for voiceover
Version 1.25.0 is almost invisible from the outside, but without it there would be no “scenes to a voice”, no voiceover and no video with no speech. Until now, in the core a sound cut and a picture cut were the same point: wherever the voice was cut, so was the video. Running the picture separately from the voice was impossible. Now sound and picture are two separate tracks, so the picture can change somewhere other than where the voice is cut. In an ordinary talking-head video they match exactly, and nothing has changed there: a run on earlier projects matched the old version byte for byte, finished video included. Face checks and the virtual camera follow the picture; speech checks, volume levelling and subtitles follow the sound.
The mode itself grew out of practice. We first edited three videos without a talking head by hand. In some, voice messages played over mute wide shots, and that meant laying in black video and stitching two projects together by hand. In another there was no speech at all: text, music and shots with the colour matched to each piece. Those videos produced two recipes, and the plan for the mode went through three reviews by different models. The reviews’ main finding was that very shared cut point between sound and picture.
Voiceover on top of your own shots
Since version 1.29.0 you can lay a voice note from your phone over your own shots. Your recording becomes the video’s sound (ogg, opus, m4a, wav and mp3 all work), and the picture is mute wide shots, other clips and photos. Before, a video like this meant putting black video under the voice and stitching two projects together by hand. Now it’s one edit and one command that gives you the finished file. The plugin places the shots under the voice itself. We tested the mode on a video made of two voice notes and 19 wide shots: the cuts and frames matched the hand-made edit, and so did the sound.
The parts can be mixed. First you speak on camera, then a voiceover plays over the shots, then you again. And the video can end with text on a background after the voice has finished.
A shot can be tied to a word: the picture changes exactly when you say it. If a clip is shorter than its place in the video, the plugin won’t freeze the last frame; it stops and shows how many seconds are needed and how many there are. A photo can stand still, slowly push in, pull out or drift to the side. If a shot already has movement in it, the virtual camera stays off on it.
Colour is set separately for each location: an office and a sunset can’t be evened out with one white balance. If bystanders who mustn’t be shown get into the frame, they have to be flagged, and the video won’t come out until a full-screen scene covers them. The rules that protect the speaker’s face (the limit on inserts, the ban on two full-screen pieces in a row) only apply where the person is actually speaking on camera. On the voiceover part they’re off: there the shots are the picture. Reading time for text and the brand tone are still checked, though.
The agent offers a video without filming on its own
You don’t have to dig through the docs to find the new modes. Since version 1.29.1, if you ask for a video without filming, the agent asks you one question and offers three options: scenes over a voice note from your phone, your voice over mute shots or photos, or a promo made of scenes only. To get started, “make a video from my voice note” is enough. If the online add-on is connected, the agent also names one more way: a voiceover in your cloned voice.
The agent also suggests the mascot unprompted, but only where it fits: in a video with no one speaking on screen, and for a brand that doesn’t have a mascot yet. In the list it’s marked as new, either drawn in code in the brand colours or your own ready-made character. With the premium tone the agent doesn’t offer it, and without your choice the mascot never goes into the video.
On the first run the plugin says in three lines what it can do, the no-camera videos included. And if you’ve updated across several versions at once, the “What’s new” block shows the headline of each one instead of the first few items of a list.
A video with no speech: text on screen, sounds and music
Since version 1.30.0 you don’t need a voice at all. The picture is your mute shots, or a person on screen who doesn’t speak, and the meaning is carried by text on screen, word for word as in the brief. The brief is split into blocks by line: hook, label, translation, punchline. How long each block stays up, the plugin works out from reading time rather than by eye, and it lays the clips under the blocks itself. If a word in the text has changed, a block is missing or shown twice, the check won’t let the video through. On real footage the length by reading time came out at 14.43 seconds, against 14.5 for the same video edited by hand.
A video with no speech can end on its punchline, held on screen while the viewer reacts, with no call to action. If a clip is slowed down below ×0.75, the plugin warns you: the frames will start repeating.
For sound, such a brief always asks. Effects go on the events in the frame, and for music there are three options: build it in code, use your own licensed track, or pick a track in the app when publishing. In that last case the video comes out with its sounds, and the report reminds you to pick the track. A licensed track that ends before the video won’t be accepted at mastering.
Since version 1.37.0 the cover of a video with no speech is the frame with the hook, the video’s main line, rather than, say, the rubric label.
For regular series there are now series marks: a rubric label at the top of the frame and the corner logo only in the last seconds. The series preset keeps them along with the voice and the music.
Music built in code
Together with the video with no speech, the core got its own music generator. It’s synthesis in code, with no neural networks and no stock libraries. The track comes out exactly as long as the video, end card included. There are four styles, which differ in instruments, rhythm and density: a steady groove, pulse, tension and ambient with no marked beat.
The music follows the text. On a turn in the story it cuts out like a stopped tape, comes back lighter under the reading, and on the punchline an accent hits, followed by silence for the reaction. Pauses, stops and accents sit exactly on the beats. Scenes can be placed on beats too, but reading time always beats the musical grid. With the same settings the track always comes out the same, and if you want something else, the plugin builds a different arrangement.
Mastering now holds −14 LUFS after the sound effects too: before, with music, the loudness crept up to −12.3 after the effects. Mastering won’t accept music built for another length or another video, and it doesn’t mistake the track’s intended pauses for the music running out.
Videos made from photos
Since version 1.35.0 photos under a voiceover or in a video with no speech have become full-fledged material. For each photo you can name the subject: an area of the picture or “faces”. A push-in then ends on the subject with room around it, as deep as the photo’s pixels allow but no more than twofold, a pull-out starts from it, a still frame centres it, and a pan travels along a group of people or a panoramic shot. The whole move is checked frame by frame in advance: if the subject can’t stay in the frame, the plugin stops, shows the numbers and suggests a way out: a pan, the whole photo or a smaller box. Scenes over a photo also frame the subject, not a default face.
Since version 1.37.0 iPhone photos in HEIC work too: you no longer need to convert them to JPEG by hand. This needs a recent ffmpeg; with an old one the plugin tells you plainly to export the photo as JPEG, and a broken frame never gets into the video.
The mode was run on seven frames from real cutaway footage saved as photos: two of them were stored sideways, one was small. From them we built a voiceover video of 16.5 seconds and a video with no speech and music built in code of 13.3 seconds. Finished videos without the new photo features matched the previous version byte for byte.
Checking the footage before editing
Since version 1.38.0, before a video without a talking head the plugin looks through your clips and points out where there may be problems:
- the camera shakes;
- there’s a tail from a video editing app left at the end of a clip;
- there are dark or overexposed stretches, black or frozen frames;
- the person in the frame is saying something, and their lips won’t match the voiceover;
- someone turned to face the camera;
- there are more faces in the frame than usual, possibly bystanders.
These are hints, not verdicts: the plugin doesn’t throw anything out or change anything in the project, the decision is yours. And what it couldn’t judge, say a sky with no detail or faces that are too small, it marks as such.
A voiceover in your own cloned voice
There are now three ways to handle the voice: record a voice note on your phone, make a video with no speech, or voice the script with a clone of your own voice. The third way exists only in the it-reelsmaker-online add-on, since version 1.6.0, through ElevenLabs, and the core still has no voice synthesis. The add-on reads the reel’s script in your voice, and from there the core edits the recording just like a voice note from your phone.
- Only your own voice: picked in your ElevenLabs account or cloned from a fresh clean recording, with your consent confirmed.
- The price first, then the “yes”: the add-on shows how many credits it will take and starts the voiceover only after you agree. If the text, the voice or the settings change, it needs a new “yes”.
- Money is never charged twice, even if the connection drops in the middle of the voiceover. And if you stop the job halfway, the add-on voices only the missing pieces.
For reference: voicing a 27-second reel cost 147 credits. At ElevenLabs prices as of November 2026, the Starter plan costs $6 a month and gives 30,000 credits, which is about 200 such videos a month. The service’s plans change, so treat this as an estimate.
How to install and update IT Reelsmaker
The plugin is still in the official Claude directory: according to the directory’s statistics for November 2026, it has more than 16,000 installs and more than 5,000 accounts active in the last 28 days. But the directory no longer accepts new versions that contain code, so its users are left with the early version 1.6.1. We went through that story in detail in the article on why the directory only takes code from Verified publishers. Editing is impossible without local scripts, so we no longer tailor the plugin to the directory. Everything covered in both articles installs from our marketplace on GitHub:
claude plugin marketplace add https://github.com/itpartypattaya/reels-montage.git
claude plugin install it-reelsmaker@itpartyIf the plugin is already installed from the marketplace, updating takes two commands, after which the skill shows a “What’s new” note once:
claude plugin marketplace update itparty
claude plugin update it-reelsmaker@itpartyIf you’re updating from a version before 1.22, two changes are worth knowing about. The plugin no longer reads environment variables: if you set the project folder or the path to the face model through REELS_PROJECT or REELS_FACE_MODEL, move them to it-reelsmaker.json or to the project settings, reel-defaults.json. And commands have been renamed: export is now stage, and set is now change. The agent knows the new names itself, but if you have your own notes or scripts with the old commands, they need fixing.
In the it-reelsmaker-online add-on, cloud transcription now accepts long recordings. The service has a limit of 25 MB per file, which is about 13 minutes of audio in the form the core prepares it, so a longer recording is sent as compressed mono mp3: a 34-minute talk weighs 8.3 MB. Word timings don’t change because of this, and the mp3 is only created after you’ve agreed to the price and is deleted after the request.
The repository’s main page has been rewritten too: it’s now a showcase, with what comes out of a single raw take, a quick start in two commands and seven steps with diagrams. A README in Chinese has appeared alongside it.
Open source and feedback
IT Reelsmaker is still released under the MIT licence; the source code and all the scripts are in the GitHub repository, and what changed in each version is in the CHANGELOG. If you’ve made your first explainer from a voice note, set up a mascot or found a bug, write to Issues: that’s exactly the feedback we use to choose what to do next.
