spoken-captions
Skill1 starsChanged 10 days agoMITView on GitHub
Put spoken words on screen from the transcript - word-timed captions on the user's own talking footage, and on a cut reel, captions per clip and a few highlight phrases timed to the words they call out, placed with scripts/reel_text.py. Use when a video is watched muted and its speech must be read, or when a short's key lines should land on screen.
1 other project uses a version of this, 1 of them edited.
Adopted by
Most recently taken up by
- PicsArt/gen-ai-skillsedited, first seen Nov 1, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 1
- PicsArt/gen-ai-skills★ 798% same
Wrote their own version · 0
None found.
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository48 other files it gives its agents.
- animate-designSkill
- app-ui-animationSkill
- asset-auditSkill
- caption-decorationsSkill
- character-riggingSkill
- custom-transitionsSkill
- design-layer-importSkill
- edit-footageSkill
- edit-on-your-computerSkill
- edit-pipelineSkill
- face-hidingSkill
- figma-storyboard-animationSkill
- film-cuttingSkill
- film-finishingSkill
- footage-analysisSkill
- footage-reframingSkill
- get-startedSkill
- large-scene-assemblySkill
- linked-element-motionSkill
- make-cartoonSkill
- make-screen-videosSkill
- motion-design-craftSkill
- motion-intentSkill
- motion-pattern-librarySkill
- motion-reviewSkill
- motion-scene-continuitySkill
- motion-tracking-from-videoSkill
- photo-slideshowSkill
- picsart-video-rulesSkill
- quality-reviewSkill
- quick-merge-and-exportSkill
- realistic-color-gradeSkill
- recreate-videoSkill
- reel-structureSkill
- reference-fragment-matchingSkill
- resting-frame-rulesSkill
- retention-hooksSkill
- scene-3d-authoringSkill
- scene-animation-referenceSkill
- scene-building-basicsSkill
- scene-template-remixSkill
- screen-recording-analysisSkill
- screen-recording-layoutsSkill
- screen-recording-zoomSkill
- screen-ui-redesignSkill
- social-postingSkill
- speech-aware-cuttingSkill
- visual-styleSkill
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 1 is distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 9
Every word on screen comes from a transcript, with its times as the transcript gives them.
- Very specific: line 15
Probe before you transcribe: Check hasAudio from picsart_media_probe first, and skip only when it reports hasAudio: false, as the …
- Very specific: line 26
Place each line as a captions layer through picsart_scene_patch: content: { kind: "captions", transcript, timeOffset }, where transcript is …
- Very specific: line 45
Read the effect's id and params from picsart_scene_capabilities (effects) rather than from here.
- Very specific: line 79
Highlights: all go on one captions layer, one segment per phrase, each on a pill the renderer fits to the measured words.
- Very specific: line 85
When the record's format sets the captions as word cards, or a hook card, ask the script for them instead of captions:
- Very specific: line 127
Check the result as the quality-review skill says, with a frame at each phrase's start.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Captions
Every word on screen comes from a transcript, with its times as the transcript gives them. Never rewrite the word times. A caption carries only what is said: never a speaker's name the viewer has not heard, a stage direction or a row label.
The transcript
- Probe before you transcribe. Check
hasAudiofrompicsart_media_probefirst, and skip only when it reportshasAudio: false, as the footage-analysis skill says. - Transcribe with
picsart_media_transcribe:- the user's own footage: each source once, as the footage-analysis skill says. Reuse that transcript; never transcribe a cut.
The tool's notes name any segment it returned without word times. Caption that segment as one
block over its own start and end; it cannot carry a highlight.
Captions on a voice-over
Place each line as a captions layer through picsart_scene_patch:
content: { kind: "captions", transcript, timeOffset }, where transcript is the tool's
transcript as returned (it is already shaped for a captions layer) and timeOffset is the
line's start in the video.
2–4 words per caption. The caption style's offset is measured from the frame center
([0, -670] puts it at y ≈ 290 in a 1080×1920 frame, below the top 250 px the platforms cover); keep captions in a dark band of the
layout, never over a panel or the focus point.
Readable over any picture, and the spoken word lit. A caption crosses moving footage, so give its style an outline and light the word being spoken:
"style": {
"effects": [{"id": "stroke", "params": {"width": "<px>", "color": "<outline colour>"}}],
"decorations": [{"target": "highlighted", "place": "text", "color": "<highlight colour>"}]
}Read the effect's id and params from picsart_scene_capabilities (effects) rather than
from here; this shape passed picsart_scene_validate on 2026-10-05. One highlight colour
for the whole video, the brand's when it has one.
Words on a cut reel
A reel cut from the user's footage plays each source in pieces, so its words are placed from the
scene the user approved, with python3 scripts/reel_text.py words.json:
{
"scene": "scene.json",
"transcripts": {"talk": "transcript-talk.json"},
"captions": {"style": {"font": {"size": 64}, "offset": [0, 440]}},
"highlights": [
{"clip": 0, "say": "revenue tripled", "text": "Revenue tripled"},
{"clip": 4, "say": "nobody tells you this"}
]
}transcripts maps each spoken source's asset id to its transcript (its file, where you run
commands). The script builds the words into the scene: where you cannot run commands, stop as
picsart-video-rules says for a script that builds the scene. Ask for captions,
highlights or both. The script reads each clip's place on the reel and its trim from the scene,
so a clip the user retrimmed in the editor is read as they left it, and it prints the scene with
the new layers on top. A phrase it had to cut short comes back in notes.
- Captions become one
captionslayer per spoken clip, open while that clip plays its source, with the source's transcript as returned. A clip held past itstrim.outfreezes on its last frame, so its captions close where the trim ends. The reel ends at the composition'sduration, or with none at the track layer's: captions stop there. A photo, amutedclip, a shot without a transcript or a clip past the reel's end gets none, and a highlight on a muted clip or one past the end is refused. - Highlights all go on one
captionslayer, one segment per phrase, each on a pill the renderer fits to the measured words. PasshighlightStyleto change the look; the font follows the scene-building-basics skill.
Word cards and the hook (a reference's format)
When the record's format sets the captions as word cards, or a hook card, ask the script for
them instead of captions:
{
"scene": "scene.json",
"transcripts": {"talk": "transcript-talk.json"},
"cards": {"words": 3, "seam": 960, "style": {"font": {"family": "bebas-neue-regular", "size": 100}}},
"hook": {"lines": ["HE'S PARALYZED.", "AND HE WON."], "box": "#C8161D", "size": 96,
"seam": 960, "gap": 70, "length": 3.2}
}- Cards: one
captionslayer of 1 to 3 words each (words), a card closing early on a word ending in. , ? !. The words keep their spoken times; each card holds toward the next, so they never blink.seamputs the cards on the seam of a split. On a split, the words come from the track with the sound (the top one); the muted bottom track is never read. A segment without word times is one card over its own span, and the last card stops where the speech does, never over the end card. - Hook: a colour box that hugs the lines (
hook-box) and the lines on it (hook-text), from the reel's start forlengthseconds, the box's bottomgappx above the seam. The lines are the ones the map approved, never written here. More than three lines, or a line over 22 characters, comes back innotes.
The style, the box colour, the size and the length come from the format, never from taste.
Choosing highlights
Highlights are a few lines worth reading, not subtitles.
- 2–4 per reel, 2–6 words each, verbatim from the transcript.
sayis what was said;textmay drop a filler word or fix its case, never change what it says. - Pick the hook, the number, the rehook (the turn near the middle, as the retention-hooks skill names it) and the payoff: the words a viewer should remember. Skip greetings, filler and anything already said on screen.
- Draft them before you place them: each phrase with its clip, so the use case can offer them. The reel is complete without them.
- One at a time. The script holds each from its first word to half a second past its last, 1.5–3 s, and never past its clip or into the next phrase.
- Big, high-contrast, inside the platform's safe band and off the face. The use case names the band. With captions on too, keep the two in different bands.
Check the result as the quality-review skill says, with a frame at each phrase's start.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

