atultiwari/youtube-video-production

Making a video, end to end

A runbook: "what do I do next", in order, with the failure behind each rule so you can argue with it rather than obey it. The order is the same for every subject; the mode changes.

0 · Choose the mode

Decide before anything else, and record it as mode: in the video's STATUS.md. It changes where reference comes from and what the picture is made of.

	Screen tutorial tutorial	Illustrated explainer illustrated	Hybrid hybrid	Motion explainer motion
Teaches	how to use an app or tool	a concept: pathology, physiology, any science	a concept, then the tool that applies it	a news event, incident or case and its lessons
Step 3	reconnaissance capture of the app	reference research: atlases, micrographs, real objects	both, per beat	a sourced fact sheet (every on-screen number cited)
Step 9	definitive capture	none	for the app beats only	none
Picture	app shots, built from captured reference	drawn pictures, through remotion-whiteboard	each beat in its own kind	typography, maps, newspaper pages, gauges, through remotion-motion-explainer

motion replaces nothing: anything drawn (a pen, a hand, stroke-by-stroke pictures) is still illustrated and still goes through remotion-whiteboard. Gate 2 in motion is the owner approving the silent preview with temporary captions.

Choose by what the viewer must see, not by habit. If no beat needs a real interface, it is illustrated. Without modes, a pathology baseline was sent to capture an app it never showed.

The order, and why it is this order

 0  Choose the mode           tutorial, illustrated or hybrid; and the style
 1  Decide the piece          angle, audience, one idea
 2  Validate the claims       before a word is scripted
 3  Reference                 capture the app, or research what to draw
 4  Script                    then lint it
 5  ── GATE 1 ──              a human reads the script before a take is spent
 5a Storyboard                illustrated: one picture per beat, events on cue words
 5b Key frames ── GATE 2 ──   the owner approves the look, on stills
 6  Record the voiceover      sarvam-tts.py: the cloned voice, one take per beat
 7  Master                    to the delivery target
 8  Measure → grid → words    the takes decide every timing, to the word
 9  Definitive capture        tutorial and hybrid: the exact flows the script calls for
10  Build                     the kit (illustrated) or app shots (tutorial), on the grid
10b Sound                     effects from the picture, a bed, ducked under the voice
11  Visual QA                 probe, one focus, fill, event gaps, a 2-s sheet, a mute review
12  Render
13  Publish                   metadata, chapters, upload, schedule

Voice before picture. The take is the clock: every beat, animation and cut is placed from the measured audio, then from each word's transcribed time. A word-count estimate was 25% wrong on a whole script, and a median 1.3 s off per word. See references/mistakes.md.

Capture twice (tutorial, hybrid): reconnaissance before the script, so it promises only what the app does; the definitive recording after, when the flow is known.

One focus at a time. One place drawn or written at a time, in reading order: the note, the picture, its label, then the pointer. The video must teach with the sound off: focus-check.py measures it, a fresh agent watching it muted tests it (09-motion.md, 11-visual-qa.md).

Two human gates. Gate 1, the script: cheap to change, takes are not. Gate 2, the look, on key-frame stills before every beat is built: without it, v2 made type on an empty dark ground.

Know where you are first: scripts/video-status.py on the video's STATUS.md (status-and-resuming.md); update the file last thing before stopping.

Non-negotiables

1. A claim gets a primary source before it gets a sentence. Secondary sources are wrong often enough that this is not optional.
2. Simulated interfaces are built from captured reference, never from memory (tutorial, hybrid). Three rebuilds of one app's composer were wrong until the real thing was measured.
3. Invented output is labelled on screen (Simulated — illustrative only); schematics say Simplified picture.
4. Nothing private appears in any frame: not in a demo, a sidebar, live-account reference, or a slide label (12-pathology-visuals.md).
5. Read written form fields back before believing them. Typed values fail silently.
6. Never hand-edit a timing file (beats.json, lines.json). They are generated.

Skills it orchestrates

Skill or kit	Its job in a video
this skill's kit	node <skill>/scripts/install-kit.mjs <project> --style paper-notes: components, style, fonts, QA and build tools. Installs remotion-whiteboard's toolkit first.
remotion-whiteboard	Required for anything drawn or hand-written: strokes, written text, the pen.
motion-design	easing and choreography for elements that are not drawn
remotion-best-practices	Remotion API usage
remotion-motion-explainer	motion mode: its own template, components, voice/caption/sound scripts and worked example (fridge-hack)
video-promo-kit	step 13 for every mode: thumbnail variants (owner picks), 1:1 carousels + LinkedIn PDF, YouTube description/chapters/tags, LinkedIn/Instagram/WhatsApp copy
youtube-uploader	step 13: the upload itself through the Data API (dry run, then --confirm on the owner's yes), thumbnail, captions, playlist, scheduling

If remotion-whiteboard is not available, stop at the build step and say so; never improvise a drawing component. install-kit.mjs enforces it.

Where the detail lives: read each file at its stage, not before

File	Stage
references/01-plan-and-validate.md	1–2
references/02-capture-reference.md	3 and 9: capture, or reference research when illustrated
references/03-script.md	4–5, including the channel asks
references/07-art-direction.md	0 (the style) and 5a–10: the house style, as data
references/08-storyboard.md	5a–5b, Gate 2
references/04-voice.md	6–8
references/09-motion.md	8 (word timing) and 10: arrivals, events every 2–4 s
references/05-build.md	10, tutorial shots, the whiteboard hand, and the drawing library (objects.tsx)
references/10-sound.md	10b
references/11-visual-qa.md, references/rubric.md	11
references/12-pathology-visuals.md	3 and 10, for medical pictures
references/13-longform.md	videos over 12 minutes, chapters
references/06-publish.md	12–13: render, ready-to-upload/ → youtube-uploader → _uploaded/ (and video-promo-kit for promo material)
references/14-lessons-2026-10.md	any mode: measured voice numbers, Whisper take check, word cues, component-fired sound, QA rules from the fridge-hack build (E01's lessons are folded into 04, 05, 06, 09, 11 and mistakes.md)
references/status-and-resuming.md	resuming a video, or recording a decision
references/mistakes.md	when something behaves oddly: read before debugging

Scripts that ship with it: each takes --help, and reads only the project it is given

Stage	Scripts
status	video-status.py (which step a video is on, by mode)
4–5a	check-script.py (TTS traps), check-storyboard.py (picture per beat, cue words, gaps)
5b	draft-takes.py (a draft grid before the takes exist, for key frames only)
6	sarvam-tts.py (takes in the cloned voice through the API; key in SARVAM_API_KEY; --respell a series map of spoken forms)
7–8	check-takes.py (Whisper every take vs the script), master-audio.py, build-audio-grid.py, word-timestamps.py (exact word times)
10	install-kit.mjs <project> (the kit, a style, fonts, tools, the Python venv), make-paper.py; in the kit, makeBeat/HL (beatspec.tsx) and the object library (objects.tsx, still test kit-objects)
10b	sound-design.py
11	focus-check.py, frame-fill.py, event-gaps.py, gap-words.py (words spoken in each static gap), contact-sheet.py; the kit's tools/qa.mjs, tools/stills.mjs (midpoints, ends, every N s), tools/build-all.sh (stops before rendering if too static)
13	chapters.py (YouTube chapters from the grid), captions-srt.py (upload captions from the word timings)

Templates: STATUS.md, STORYBOARD.md, MUTE-REVIEW.md in templates/. Styles: styles/*.json.

Formats

Short: 1080×1920, delivered at 2160×3840 (build-all.sh --deliver-4k). 26–56 seconds, one idea; the first line and the first frame decide everything. Bottom 360 px clear for the UI.

Longform: 1920×1080 at 30fps; seven to twelve minutes, or 15–30 for an illustrated lecture built in 3–5 minute chapters (13-longform.md). Optional second language as an alternate track.

What is dated

Prices, model names, tier boundaries and statute dates in the references were correct on the measuredOn date above. Anything marked VERIFY is re-checked against its primary source before it is said on camera. Never inherit a number from this skill into a script.