An AI agent skill for helping an instructor create a coherent, accurate, editable course video on any subject.

繁體中文 · Download v1.2.0 · Skill entrypoint · Changes

The skill covers the whole teaching-production loop: understanding the audience, selecting a learning progression, checking evidence, designing explanations, synchronizing slides and speech, producing media, reviewing the finished lesson, and handling revisions and authorized delivery.

It grew from educational media production and a closely revised Interspeech conference tutorial collaboration. The result generalizes the teaching decisions and engineering lessons instead of distributing the original tutorial, private feedback, voice, or topic-specific rules.

• Organize around what learners need to understand or do, with sources serving the explanation.
• Make actors, decisions, comparisons, and timing explicit in difficult diagrams.
• Diagnose weak flow by reading every transition together and checking the opening promise against every scene.
• Coordinate spoken ideas, visible objects, and reveal/highlight events.
• Place pauses according to thinking and viewing needs, rather than applying a fixed gap everywhere.
• Verify claims with their actual conditions, denominators, source versions, and uncertainty.
• Keep editable slides, notes, scripts, takes, video, captions, and downloads aligned during revision.
• Use speech recognition as a locator: compare numbers, percent/minus signs, and negations as well as overall similarity, read the chosen take's transcript, and re-check the exported audio after post-processing.
• Measure slide changes against speech onsets in the exported file (start, middle, and end), not in a computed timeline.
• Tell viewers when narration is synthesized or voice-cloned, or realistic imagery is generated.
• Review correctness, comprehension, and technical integrity separately.

These are decision tools, not a fixed presenter persona or slide formula. Research tutorials, software demonstrations, physical-skills lessons, humanities talks, business briefings, and workshops each get a different treatment.

Download tutorial-video-skill-v1.2.0.zip from the release page, or clone this repository:

git clone https://github.com/speechlab0210/tutorial-video-skill.git

For an agent with a skill-folder mechanism, place the entire tutorial-video/ directory, including references, assets, and scripts, in its configured skills directory. The release ZIP unpacks to a tutorial-video-skill/ folder; copy only its tutorial-video/ subfolder, for example to:

• ~/.claude/skills/tutorial-video/ (Claude Code, personal) or <project>/.claude/skills/tutorial-video/ (Claude Code, one project)
• ~/.codex/skills/tutorial-video/ (Codex)

Do not copy only SKILL.md. Other agents can read that file directly and follow its linked references; host-specific automatic discovery is not guaranteed.

Example request:

Use the tutorial-video skill to create a 20-minute beginner lesson about urban gardening. The audience has no gardening experience. Start from my notes, use editable slides, and make the examples explain decisions a learner can actually make. Produce the video, captions, and transcript locally.

Existing material:

Use the tutorial-video skill to revise only section 2 of this lecture. Preserve the accepted sequence and voice. Explain the confusing diagram with one concrete example, and synchronize notes, audio, captions, and video. Leave other sections unchanged.

The skill respects draft-only, read-only, and narrow-revision requests. It honors explicit publishing or email authorization the user has already given for this work, destination, and visibility, but does not treat video creation as automatic permission to publish, share, or email.

Python 3.10+ is sufficient for planning, speech-transcript checks, caption structure checks, and manifests (use python3 or py -3 where python is unavailable). There are no Python package dependencies and no network calls. Run from this repository (paths are examples; speech, captions, and manifest check files you produce later: a recognizer transcript of a take or of the exported audio, an exported SRT, and a folder of finished deliverables). With an installed skill, work in your own project folder and replace tutorial-video/ below with the installed skill folder's full path:

python tutorial-video/scripts/project.py init my-lesson
python tutorial-video/scripts/project.py check tutorial-video/assets/examples/weighted-averages.json
python tutorial-video/scripts/project.py outline tutorial-video/assets/examples/weighted-averages.json
python tutorial-video/scripts/project.py speech --project my-lesson/project.json --scene S1 --heard my-lesson/asr/S1.txt
python tutorial-video/scripts/project.py captions my-lesson/captions.srt --duration 120
python tutorial-video/scripts/project.py manifest my-lesson/deliverables --out my-lesson/manifest.json

The initialized plan is deliberately incomplete. Fill it before checking it. A structural pass is not an educational quality rating.

speech compares a recognizer's plain-text transcript (from any recognizer you are authorized to use; the helper runs none) with the scene's narration and transition, or with a text file via --script. It reports similarity, missing/extra/replaced passages, the numbers on each side with written and spoken forms treated as equal (1600, 一千六百, one thousand six hundred; 5萬7千; 6:09 = 六點零九分), percent and minus signs, and negation words. It exits 1 when numbers, signs, or negations differ, or when similarity falls below a --min-similarity you calibrated. It does not listen: the differences are places to hear, and misreadings that the recognizer silently corrects stay invisible to it. Calibrated on 1,397 accepted Chinese and 103 English narration segments, it flagged a number on 4.4% and 2.9% of them, nearly all recognizer homophones, while catching every planted number and negation error.

For narrated still scenes, install FFmpeg and FFprobe with H.264/AAC support, prepare real images and authorized audio, copy tutorial-video/assets/templates/render.json into your project (e.g. my-lesson/), edit its asset paths, and run:

python tutorial-video/scripts/assemble.py my-lesson/render.json --out my-lesson/video.mp4

This produces a new movie and a timeline receipt. Narration is encoded once across the whole timeline and every scene starts on the frame grid, so speech stays on its slide however many scenes there are; the assembler checks decoded audio length against video length before writing the file. It does not synthesize speech, create slides, normalize loudness, generate captions, or inspect comprehension. Animated slides and live demos need an appropriate editor. See the contracts and limits.

python -m unittest discover -s tests -v

Tests exercise malformed plans, coverage and ID errors, speech-transcript checks (Chinese and English number forms, changed numbers, ranges, dropped percent and minus signs, dropped negations, digit-by-digit readings, substitutions, script-variant warning, input validation, and an exit code per check), Unicode captions and console output, overlap and duration errors, hash integrity, link rejection, no-overwrite behavior, and rendering path boundaries. Media integration tests build synthetic movies and measure the shipped MP4: scene order and retained ending silence; a 12-scene sync test at 30 and 7 fps that compares every slide change with the matching speech onset; a VBR MP3 whose container under-reports its length; awkward slide files (a % in the name, transparency, GIF, a JPEG with a .png name); and error messages that name the failing scene. To run them, put FFmpeg/FFprobe on PATH or set FFMPEG_BINARY and FFPROBE_BINARY to their full paths (only the tests read these variables). Run with -v and read any skip reasons: "FFmpeg/FFprobe unavailable" or "libmp3lame unavailable" means media integration was not verified; "cannot create a symlink here" (or junction) only means the account cannot create that kind of link.

The repository also includes manual behavioral evaluation cases. Passing automated tests proves these helper behaviors, not cross-agent compatibility, factual accuracy in arbitrary subjects, or learning effectiveness. No controlled learner study is claimed.

This package was written by an AI agent and is maintained through the speechlab0210 account. It synthesizes reusable experience; it is not an official Interspeech resource or a reviewed endorsement by an instructor. It includes no private teaching correspondence or original conference media.

The original instructions, templates, examples, and code in this package are provided under the MIT license. Third-party tools and material used in future lessons retain their own licenses and permissions.