Personal Brand & Creator Video

THE LOOK IS CHOSEN, NOT MANDATED. Pick it in the Style Pass — a motion style + a UI style (MOTION-UI-STYLE-LIBRARY.md) and a caption style (CAPTION-STYLES.md) — chosen for THIS job and recorded in CONCEPT.md. The Apple Standard (section 3) is the house default and a strong starting point for most product, UI and corporate work, but it is a recommendation, not a rule. If another style fits the brief better, recommend it with your reasoning and use it. No style is deprecated.

Scope. Vlogs, tutorials, opinion/commentary, day-in-the-life, channel content, founder-led content. Authenticity is the craft; the person is the asset.

Asset tier: 2 — MANDATORY (the person). The creator's footage is required — we cannot invent a person. Only a graphics-only "concept explainer" can be built from nothing; a personal-brand video needs the actual human.

STEP 0 — SOURCE GATE (mandatory, BEFORE the assets prompt)

You cannot plan assets without the material. Before writing ASSETS-PROMPT.md, obtain at least ONE of:

the video clip — the footage to edit, or
the voiceover / audio — the narration track, or
the transcript / script — the words.

Ask for it up front. If the build is genuinely from-scratch graphics (no source exists), say so and record it — the concept you write becomes the script.

Then analyse the source and save the evidence as SOURCE-ANALYSIS.json:

If a clip -> video analytics: ffprobe (codec, size, fps, duration, audio channels); scene detection for cut times and ASL (ffmpeg -i in.mp4 -filter:v "select='gt(scene,0.3)',showinfo" -f null -); loudness (integrated LUFS + true peak); palette sample (quantise frames -> hex); beat/BPM if there is music.
If audio -> word-level transcription: faster-whisper with word_timestamps=True -> a JSON of {word, start, end} per word, plus segments. This drives the Visual Narration Plan, the captions and the word sync. Recipe: decode to 16 kHz mono WAV with ffmpeg, pass a numpy array, run with MKL_THREADING_LAYER=GNU OMP_NUM_THREADS=1 and cpu_threads=1, model tiny (base/small for accuracy).
If text only -> the sentence list, mapped to the Sentence Law table.

Only once the source is in hand and analysed do you write ASSETS-PROMPT.md — so every prompt reflects what the build actually needs.

REFERENCE-LEARNED PATTERNS (from captured presets)

Real-world patterns measured from reference edits in this vertical. Apply them; the matching preset locks the exact look (see presets/INDEX.md).

From presets/creator/preset-007-creator-kinetic-text (a 76 s creator talking-head, "infotainment"):

Kinetic word-by-word text overlaid on the chest, synced to speech (~every 0.5 s) — the text IS the edit.
One accent colour (orange) for a script connective + pill badges; everything else white/black.
Outlined grey section numerals (01 / 02 / 03) to mark structure.
Screenshot / UI proof cutaways (dark-mode UI, phone mockups, DMs).
A colour-inverted emphasis frame for a hard beat.
A 3-step structure (hook / body / CTA) made explicit on screen.
Fast: ~2 s ASL.

From presets/short-form/preset-002-realtor-word-caption: word-by-word captions with a two-colour keyword accent system; a strong hook ("STOP SCROLLING"); a frosted- glass pill end card.

STEP 0.5 — A-ROLL PREP (matting first — mandatory when a person speaks)

If this piece has a person speaking to camera (talking-head, voiceover, avatar, podcast), the first job is the background, before the concept. Decide the path in a-roll-matting: keep it / matte it / key it — and by default get the character off the background (or onto green). Matte FIRST unlocks text-behind-subject, screen replacement, graphic backgrounds and floating UI. Use tools/matte.py for the local matte (rembg + ffmpeg). Record the matte as an asset in the manifest and check it at the QA gate (no holes, no baked- caption artifacts, stable alpha). If the background is the message, keep it.

STEP 1 — CONCEPT.md (write the plan before any prompt)

With the source analysed, write CONCEPT.md — the single plan the whole build follows. Everything is connected: every line here drives a later stage, and every later stage writes its result back here. Follow THINKING-SYSTEM.md — the planning stack, the four lenses (EZRA), and (when you have only a script) the only-a-script path.

Premise — the video in one sentence + its emotional arc.
THINKING PASS (mandatory). Work the stack top-down — goal -> audience -> angle -> concept -> script -> beats -> shots — and record:
- the target emotion per section (EZRA: Emotion, Story, Rhythm, Action);
- the beat map with the two-column (said | shown) filled for every row;
- the visual plan table: beat -> viewer question -> visual evidence -> risk to review -> final asset;
- the retention check — where the video is most likely to lose people, and the fix.
Style — the named build style (this skill) and how it applies here.
STYLE PASS (mandatory). Pick deliberately and record: the motion style and the UI style from MOTION-UI-STYLE-LIBRARY.md, and the caption style from CAPTION-STYLES.md. One primary + at most one garnish; name the chosen style's failure mode.
Segment plan — the beat map with timings, taken from the analysis.
Sentence table — one row per narration sentence: sentence -> visual concept -> lane (A speaker / B visual) -> the stressed word to land on -> timing -> element bindings -> camera (reason + target + zoom). (The Sentence Law.)
Camera-track plan — the ordered camera entries (READ / EMPHASIZE / REVEAL / FOLLOW / BREATHE) with their word bindings. (The Camera Law.)
FEATURE MAP — the Feature Pass (mandatory). Walk the full catalogue in ADVANCED-FEATURE-USE-CASES.md — camera & framing, motion & animation, speed & time, transitions, text, colour, compositing & VFX, audio, AI, stills & design, workflow — and record for every feature whether it applies and how: feature -> applies? -> where (scene/timecode/sentence) -> how (implementation) -> why (the job it does). Every group is visited; no group is skipped. The concept is not finished until every applicable advanced feature has a row (a "yes" with no how is not a plan; every "no" is a deliberate choice).
Contact-sheet plan — which sign-off variants will be built (V1 Classic Grid / V2 Storyboard Filmstrip / V3 Pro QC Sheet) and why. The variant the user picks is written back here.
Sync map — the word-level timings that drive text and visuals.
Asset manifest — what the build needs (fed by the feature map); this feeds STEP 2.

Then STEP 2 writes ASSETS-PROMPT.md from this plan.

PIPELINE CONNECTIVITY LAW (mandatory)

Everything is connected — no stage is decided in isolation, and no stage is skipped:

SOURCE -> analysed into SOURCE-ANALYSIS.json, which feeds the concept.
CONCEPT -> drives the thinking pass, the style pass, the asset manifest, the camera track, the sentence table and the feature map; it is the single source of truth for the build.
ASSETS-PROMPT -> written from the concept's manifest; nothing unplanned appears in the build.
BUILD -> follows the concept's sentence table, camera track and feature map exactly.
RENDER GATE -> the contact-sheet variants visualise the concept's beat map; the variant the user picks is written back into CONCEPT.md.
QA GATE -> edit-qa-validator / tools/qa_check.py re-check that every feature the map marked "yes" actually made it into the edit.
SIGN-OFF -> RENDER -> the full render happens only after the variant AND the cut are finalised.
A new contact sheet means a new CONCEPT revision. If the sheet reveals a change, the concept is updated FIRST, then the build follows. The record stays connected end to end: source -> concept -> assets -> build -> sheet -> sign-off -> render.

ASSETS-PROMPT.md — MANDATORY (strict rule)

After the concept, produce ONE file — ASSETS-PROMPT.md — listing every asset the build needs, one executable brief per item, addressed to an AI agent that has image generation, audio generation AND coding. No prompt list, no build.

Images — image briefs: subject · composition · style · palette (hex) · lighting · aspect + background · negatives.
Transparent images (PNG/alpha) — briefs ending "transparent background, PNG with alpha, no background".
Logos — a brief, or "client supplies".
Music — audio briefs: mood · genre · BPM · length · instrumentation.
Sound effects — audio briefs: type · character · duration.
Code components — CODING briefs for what code does best (glass cards, animated type, diagrams, count-ups, UI mockups, particles, shaders, 3D): exact values (hex, px, easing, durations), the motion, and the nested-zip deliverable: ONE master zip containing a zip per category and a zip per component kit (each kit unzips to index.html, styles.css, README.md, assets/, palette exposed as CSS variables).

This file IS the prompt — write it as a self-contained instruction you hand straight to the agent, ending with the deliverable tree (one zip) and acceptance checks. A full worked example: EXAMPLE-ASSETS-PROMPT.md.

EXCLUDED — never in this file: voiceover and ALL video clips (A-roll and B-roll). The user supplies the A-roll — voice, primary footage, or a transcription JSON — at the start. If the build needs any B-roll clip, ask the user for it separately; clips are never listed in this prompt file. (The agent can only generate images, audio and code — it cannot generate video.)

Full format, the kit shape, and a worked example: ASSET-REQUEST-GUIDE.md.

1. Intake — ask before you build

What is your channel about, and who is it for?
Which format: vlog, tutorial, opinion, day-in-the-life, or talking-head explainer?
Do you have footage, or a script to read to camera? (I can cut either)
Your colours/font/logo, and any intro/outro you already use?

Then emit ASSET-REQUEST.md:

# ASSET REQUEST — <creator>
## Footage (MANDATORY)
1. Talking-head takes — 1080p+, eye-line level, separate audio if possible
2. B-roll of your day/work/process — 10s+ per shot
## Script (if you have one) or the raw takes to cut
## Brand: name, colours, font, intro/outro assets
## Music & SFX — prompts in ASSETS-PROMPT.md (your voice supplied separately)

If only audio arrives, build a talking-audiogram and say so.

2. Structure

Vlog: hook (a moment from the day) → the arc (start → middle → end) → reflection → next-video hook.
Tutorial: hook (the result) → the steps (one per beat) → the recap → CTA.
Opinion: hook (the claim) → why it's true (3 points) → the counterargument → your conclusion.
Day-in-the-life: hook (the most interesting moment) → chronological beats → a closing thought.

3. The edit — authenticity rules

Keep the person's voice and quirks; cut only filler and dead air (≤0.3s).
Jump cuts are the rhythm — mask the harshest with a zoom change or B-roll.
B-roll over every claim or story point; return to face visibly advanced.
On-screen text maps the key point (not subtitles — Text Law). One idea per beat.

3. The look — the house default (Apple Standard) + the Style Pass

The house default. The Apple Standard below is the pack's default look and the right starting point for most product, UI and corporate work — verified tokens, one accent, restrained motion. It is a recommendation, not a mandate: choose the look in the Style Pass (MOTION-UI-STYLE-LIBRARY.md). If a different style fits this brief better, recommend it with your reasoning and use it. No style is deprecated.

Canvas: alternating #ffffff / #f5f5f7 bands — the colour change IS the divider (no borders, no rules). Dark variant #0d0d0f with glow rgba(10,132,255,.25).
Text: #1d1d1f primary · #6e6e73 secondary · #86868b tertiary · hairlines #d2d2d7 on light, #333 on dark.
The one blue — interactive only, never decoration: filled action #0071e3 · text links #0066cc on light · #2997ff on dark. Hover wash #e8e8ed. Icon gradient #41A6FF -> #0A64E8 at 180°.
Type (Inter stands in for SF Pro): hero 80px/600/-1.2px · display 56px/600/-0.28px · section 40px/600 · body 17px/400/25px/-0.374px · small 14px/-0.224px · caption 12px. Negative tracking at EVERY size. One accent word per headline, never whole lines.
Buttons: pill radius 980px, filled #0071e3, white text, 44px tall, 11px 21px padding (compact 36px/14px); outline twin 1px #0066cc. One CTA per scene, maximum.
Shadow: one light source. Cards 0 24px 60px rgba(0,0,0,.08) + hairline rgba(0,0,0,.055); icons 0 30px 70px rgba(10,100,232,.35) + inner highlight inset 0 2px 6px rgba(255,255,255,.45). Shadow on the focal element only.
Radii: cards 28-32px · icons ~26% of size · pills 980px · inner UI 12-16px · never below 10px. Spacing on the 8px grid; card padding 34-40px.
Frosted bar (nav / chapter strip): rgba(250,250,252,.8) + backdrop blur, 44px tall, 1px hairline #d2d2d7 beneath.
Motion: springs with 5-8% overshoot; every entrance = fade + rise + scale-settle (0.5->1) + deblur 12-18px->0; stagger 0.05-0.16s; exits scale to ~1.05 with blur 8; camera push-ins <=8% over 3-5s. Never linear.
Layout: symmetric, centred, one focal point, >=30% whitespace, <=6 elements per scene.

Vertical application for this skill: clean Apple type over the footage, one accent for the keyword, Apple cards for the lower-thirds. Consistent, never busy.

5. Audio

Your voice highest and clean (de-noise, −16 LUFS); lo-fi bed far back (60–90 BPM); SFX sparingly. Music must be licensed.

6. Delivery

16:9 for YouTube long-form; 9:16 for Shorts/Reels (hook ≤1s). Chapters for long videos. −14/−16 LUFS.

Visual narration layer — full spec (mandatory wherever a person speaks)

The market-dominant format: the speaker carries the voice, the visuals carry the meaning. Whoever is speaking — on camera, walking, or voiceover — the video must SHOW what is being said.

Three placements (choose per beat, alternate them)

Full-screen cutaway — visual takes the frame (icon, diagram, stat count-up, keyword card, B-roll). Voice continues (L-cut in, J-cut out). Best for concepts, numbers, lists, comparisons.
Front overlay — motion graphics over the speaker; face stays visible (labels, keyword pops, callouts, arrows, lower-thirds). Best for emphasis, naming, quick facts.
Behind / around the subject — speaker keyed/cut-out over graphic background or blurred plate, elements animating behind and beside. Best for intros, hero segments, brand pieces. Needs a clean cut-out or a clean plate (Tier 1–2 asset).

Rules

Every spoken concept gets a visual landing on the exact word (±100 ms).
One visual event every 6–10 s; no talking frame static > ~8 s.
Alternate placements; avoid 3 of the same in a row.
Captions are additive, never the layer — remove them and the ideas must still show.
One idea per visual. Motion: fade + rise + settle, eased, never linear (180–450 ms in, 250–350 ms out).
Never hide the speaker's face when the face is the message.
Zero-speech pieces: apply the same layer to on-screen kinetic text.

The Visual Narration Plan (write BEFORE cutting)

Beat micro-timings

Keyword pop lands 0–80 ms before the audible word onset (the eye leads the ear).
Count-ups run 0.6–1.0 s and finish on the spoken number.
Lower-third: in at first speech, on screen 4–5 s, out ≥ 0.3 s before the next cut.
Cutaway length = the length of the spoken idea (usually 2–5 s); return to speaker on the sentence boundary.

Industry benchmarks (working conventions)

YouTube long-form 6–12 min with a chaptered structure; hook ≤ 10 s; pattern interrupt every 6–10 s.
Voice clean (de-noise, −16 LUFS), music far back, face never fully hidden.
Thumbnail and first 15 s are one decision: promise → proof → open loop.
Authentic texture over polish — consistent but human.

Worked example

Example — 8 min tutorial

t	Beat	Edit
0–10	Hook + promise	result clip + "By the end you'll…"
10–60	Context	talking head + 2 cutaways
1–6 min	3–4 steps	screen/B-roll, step cards, keyword pops
6–7.5	Recap	summary card
7.5–8	CTA	one next video

Common mistakes to avoid

Long slow intros.
Over-animated visuals that don't map to speech.
Inconsistent audio level between segments.

Signature techniques for this style (use these, with the recipes below)

Jump-cut talking head with alternating 100% / 115–125% framing (recipe 1) to hide cuts.
Keyword pops + word-pop captions (recipe 13) mapped to what is said.
B-roll insert on every named object, 1–3 s, slight push.
Chapter cards and a progress bar for long videos.
Voice chain (recipe 17) and a low music bed ducked (recipe 18).
Open-loop hook + pattern interrupts every 6–10 s.

Modern editing toolkit (techniques editors use today + how to do them from the terminal)

A. The technique catalogue — what top editors actually reach for

Cutting & structure — jump cut · J-cut / L-cut · match cut (shape, motion, colour) · smash cut · cutaway/insert · split edit · cut-on-beat montage · invisible cut hidden by a whip, zoom or object wipe · freeze frame · reverse · seamless loop · transcript-based rough cut · silence/filler removal · multicam switching. Motion & camera — eased punch-in/push-in · slow digital push · pan/tilt on stills (Ken Burns) · 2.5D parallax from separated layers · camera shake on impacts · speed ramp / time remap · optical-flow slow-motion · stabilisation · whip pan · simulated dolly-zoom · motion blur on fast moves. Transitions — hard cut is the default; then push/slide, whip, zoom-through, shape/mask reveal, luma wipe, light-leak/flash, glitch/RGB split, crossfade. Every transition has a matching sound. Never repeat the same transition twice in a row. Text & captions — kinetic typography · word-by-word karaoke captions · one-accent-word highlight · animated lower-thirds · text tracked to a moving object · text behind the subject (matte) · typewriter/mask reveals · drawn-on callouts and arrows · count-up numbers. Compositing & depth — subject cut-out/matting · background blur/replace · split-screen · picture-in-picture · chroma/luma key · screen replacement (UI on a device) · shadow + reflection under cut-outs · glow/bloom · overlays (grain, dust, light leaks) · blend modes (screen, add, multiply). Colour & look — correct first (exposure, white balance), then log→Rec.709, LUT, contrast curve, secondary tweaks (skin, sky), split-tone/film emulation, halation/bloom, grain, vignette, letterbox (2.39:1), shot matching. Protect skin tones. Audio design — dialogue chain (high-pass → denoise → de-ess → compress → loudnorm) · side-chain ducking · SFX layer (whoosh, impact, riser, tick, sub-drop) · ambience/room tone · music edited to phrases · a 0.2–0.4 s silence before the drop · stem separation for cleanup. Graphics & data — animated charts with a single highlight colour · count-ups · map/route draws · diagrams that draw on the narration · UI demo with eased cursor + click ripple · stat "bento" grids · progress bars · timeline graphics. AI-assisted (2026) — word-level transcription · filler/silence removal · scene detection · auto-chapters · face-tracked auto-reframe · matting/rotoscope · upscaling · frame interpolation · stem separation · beat detection · generative B-roll/imagery (disclose when real footage is implied).

B. Terminal toolchain (pick the lightest tool that does the job)

Tool	Use it for	Notes
FFmpeg / ffprobe	cuts, concat, scale/crop, zoom, xfade transitions, speed ramps, LUT/grade, grain/glow, overlays, captions (ASS), audio chain, loudness, export	The backbone. Chain edits in ONE `filter_complex` pass to avoid generation loss.
auto-editor	rough cut by removing silence/dead air via loudness/motion analysis	Signal analysis, not generative.
faster-whisper / WhisperX	transcription with word-level timestamps (captions, beat plans, filler cuts)	Needed for ±100 ms word sync.
PySceneDetect / ffmpeg `select='gt(scene,0.3)'`	find shot boundaries in source footage
librosa (Python)	BPM + beat timestamps for cut-on-beat
HyperFrames (HeyGen, Apache-2.0)	motion graphics as HTML/CSS/GSAP/Lottie/Three.js → deterministic MP4; CLI `init`, `preview`, `lint`, `render`	Agent-friendly: `npx hyperframes init my-video` → edit `index.html` → `npx hyperframes render`.
Remotion (React)	code-defined video compositions, data-driven templates	Check its licence terms for company/commercial use.
MoviePy (Python)	scripted clip assembly where FFmpeg graphs get unwieldy	Slower than raw FFmpeg.
rembg / Robust Video Matting / SAM-family	subject cut-out (alpha) for text-behind-subject and layered looks	Check each model's licence (some are non-commercial). Review edges/hair; rembg is per-frame (flicker risk), RVM is temporally consistent.
MediaPipe	face tracking for auto-reframe to 9:16
Demucs	separate vocals/music for cleanup
Blender (headless `blender -b -P script.py`)	3D titles, product/architecture renders, camera moves	Heavy; use only when 3D is the point.
MLT/melt, GStreamer, OpenTimelineIO	timeline-style assembly and interchange	Optional; not needed for most jobs.
ImageMagick / Pillow	stills, masks, cards, contact sheets
Always check what is installed first and fall back to FFmpeg-only: `for t in ffmpeg ffprobe python3 node npx auto-editor scenedetect melt blender; do command -v $t >/dev/null && echo "have $t"		echo "missing $t"; done`. If the network is off, do not plan on installing anything.

C. Tested FFmpeg recipe cookbook (verified on FFmpeg 6.1.1; confirm a filter exists with `ffmpeg -filters | grep <name>`)

Set IN=input.mp4. All outputs add -c:v libx264 -pix_fmt yuv420p (+ audio as needed).

# 1 Smooth punch-in (6%/s growth, centred). Prefer this over zoompan on VIDEO — zoompan can jitter.
-vf "scale=w='trunc(1920*(1+0.06*t)/2)*2':h='trunc(1080*(1+0.06*t)/2)*2':eval=frame:flags=bicubic,crop=1920:1080"
# 2 Transition + matching audio crossfade (offset = clip1_duration − transition_duration; clips need same size/fps/pixfmt)
-filter_complex "[0:v][1:v]xfade=transition=smoothleft:duration=0.5:offset=3.5[v];[0:a][1:a]acrossfade=d=0.5[a]" -map "[v]" -map "[a]"
#   other xfade names: fade fadeblack fadewhite wipeleft slideleft circleopen circleclose radial pixelize hlslice vuslice dissolve smoothup distance
# 3 Speed ramp (normal → slow-mo → fast) with split + setpts + concat
-filter_complex "[0:v]split=3[a][b][c];[a]trim=0:1,setpts=PTS-STARTPTS[v1];[b]trim=1:2,setpts=(PTS-STARTPTS)*2[v2];[c]trim=2:4,setpts=(PTS-STARTPTS)*0.5[v3];[v1][v2][v3]concat=n=3:v=1:a=0[v]" -map "[v]"
# 4 Premium look: glow/bloom + film grain + vignette
-vf "split[a][b];[b]gblur=sigma=30,eq=brightness=-0.05[g];[a][g]blend=all_mode=screen:all_opacity=0.25,noise=alls=10:allf=t,vignette=PI/5"
# 5 Colour: contrast curve + split-tone + saturation (add lut3d=file.cube first if you have a LUT)
-vf "curves=preset=medium_contrast,colorbalance=rs=-0.05:bs=0.08:rh=0.08:bh=-0.06,eq=saturation=1.08"
# 6 Camera shake for impacts (apply with enable='between(t,a,b)' on a crop wrapper, or to a short segment)
-vf "scale=2016:1134,crop=1920:1080:x='48+12*sin(t*40)':y='27+8*cos(t*47)'"
# 7 Whip-style blur around a cut at t=2.0 + slight push
-vf "boxblur=luma_radius=40:luma_power=1:enable='between(t,1.8,2.0)',scale=iw*1.1:ih*1.1,crop=1920:1080"
# 8 Cinematic 2.39:1 letterbox
-vf "crop=1920:804,pad=1920:1080:0:138:black"
# 9 Vertical 9:16 from 16:9 with blurred-fill background
-filter_complex "[0:v]split[a][b];[a]scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920,gblur=sigma=40[bg];[b]scale=1080:-2[fg];[bg][fg]overlay=(W-w)/2:(H-h)/2"
# 10 Split screen
-filter_complex "[0:v]scale=960:1080,setsar=1[l];[1:v]scale=960:1080,setsar=1[r];[l][r]hstack"
# 11 TEXT BEHIND SUBJECT: text on background, subject (alpha from matte) composited on top
-i bg.mp4 -loop 1 -i subject.png -loop 1 -i matte.png -filter_complex "[0:v]drawtext=text='PREMIUM':fontsize=420:fontcolor=white:x=(w-text_w)/2:y=(h-text_h)/2:fontfile=FONT.ttf[bg];[1:v][2:v]alphamerge[fg];[bg][fg]overlay=shortest=1"
#    for video: produce a per-frame matte (rembg/RVM) → ProRes 4444/WebM alpha subject clip → overlay it over the text layer.
# 12 Animated drawtext (fade + rise over 0.5 s)
-vf "drawtext=text='HELLO':fontsize=96:fontcolor=white:x=(w-text_w)/2:y='h/2+60*(1-min(t/0.5,1))':alpha='min(t/0.5,1)':fontfile=FONT.ttf"
# 13 Karaoke / word-pop captions: write an .ass file (per-word {\kf} timing from WhisperX, {\t(...)} scale pop) then burn
-vf "subtitles=captions.ass"
# 14 Progress bar
-vf "drawbox=x=0:y=ih-12:w='iw*t/DURATION':h=12:[email protected]:t=fill"
# 15 Freeze-frame hold of 1 s at the end / reverse / interpolate to 60 fps
-vf "tpad=stop_mode=clone:stop_duration=1"   |   -vf reverse   |   -vf "minterpolate=fps=60:mi_mode=mci"
# 16 Stabilise (2-pass)
ffmpeg -i $IN -vf vidstabdetect=result=t.trf -f null -  &&  ffmpeg -i $IN -vf vidstabtransform=input=t.trf:smoothing=15 out.mp4
# 17 Dialogue chain + loudness (measure first, then apply the measured values for true two-pass)
-af "highpass=f=80,afftdn=nf=-25,acompressor=threshold=-18dB:ratio=3:attack=15:release=200,loudnorm=I=-16:TP=-1:LRA=11"
ffmpeg -i $IN -af loudnorm=I=-14:TP=-1:LRA=11:print_format=json -f null -     # read input_i/input_tp/… then rerun with measured_* values
# 18 Music ducking under voice (voice = input 0 audio, music = input 1)
-filter_complex "[1:a][0:a]sidechaincompress=threshold=0.05:ratio=8:attack=20:release=300[m];[0:a][m]amix=inputs=2:normalize=0[a]" -map 0:v -map "[a]"
# 19 Detect silence / scene changes (feed results into a cut list)
-af silencedetect=n=-35dB:d=0.3 -f null -      |      -vf "select='gt(scene,0.3)',showinfo" -f null -
# 20 Final export (H.264, tagged Rec.709, web-ready)
-c:v libx264 -preset slow -crf 17 -pix_fmt yuv420p -profile:v high -colorspace bt709 -color_primaries bt709 -color_trc bt709 -c:a aac -b:a 256k -ar 48000 -movflags +faststart
# 21 RGB split / chromatic aberration (glitch accent — apply to a 3–6 frame window only)
-filter_complex "split=3[r][g][b];[r]lutrgb=g=0:b=0,pad=iw+8:ih:8:0[r2];[g]lutrgb=r=0:b=0,pad=iw+8:ih:4:0[g2];[b]lutrgb=r=0:g=0,pad=iw+8:ih:0:0[b2];[r2][g2]blend=all_mode=addition[rg];[rg][b2]blend=all_mode=addition,crop=1920:1080:4:0"

Rules that prevent bad renders: conform every source to the same fps/size/pixel format before xfade/concat · work from proxies (720p) while iterating, render masters once · chain filters in one pass · never mix variable-frame-rate screen recordings (convert with -vsync cfr -r 30) · keep a 1.1–1.2× scale margin before any crop-based move to avoid black edges · use -ss before -i for fast seeking, after -i for frame accuracy.

D. The production loop (every job)

Probe inputs (ffprobe -v error -show_entries stream=codec_name,width,height,r_frame_rate,duration -of json). 2. Transcribe with word timestamps. 3. Plan: write the cut list + Visual Narration Plan (timecode, phrase, placement, visual, sound) as a table or JSON. 4. Rough cut (silence/filler removal, scene boundaries). 5. Motion graphics rendered separately (HTML/HyperFrames/Remotion or FFmpeg drawtext/overlay) with alpha where they sit over footage. 6. Composite + grade + sound design. 7. Loudness + export. 8. Verify like a viewer: extract frames at key beats (ffmpeg -ss T -i out.mp4 -frames:v 1 f_T.png), make a contact sheet, check safe zones, check captions against the transcript, re-measure loudness. Fix and re-render; never ship unseen.

E. What makes it feel premium (rules, not effects)

One system: one palette, one type pair, one motion grammar, one transition vocabulary per video. Consistency reads as expensive.
Ease everything: entrances cubic-bezier(0.16,1,0.3,1), transforms cubic-bezier(0.65,0,0.35,1), playful overshoot cubic-bezier(0.34,1.56,0.64,1). Never linear.
Depth: separate foreground / mid / background; add soft shadows, subtle parallax (≤ 3–6% travel), grain 5–12% and a gentle vignette.
Hierarchy: one focal point per frame, ≥ 30% negative space, max ~6 elements.
Rhythm: vary shot length; land graphics and cuts on beats/words; let important moments breathe.
Sound sells picture: a sound for every cut and graphic land; the quiet before the hit.
Restraint: effects serve the idea. If it doesn't clarify or emphasise, delete it.
Polish: no audio pops, no jitter, no black edges, no text outside safe zones, no single-frame flashes.

Hybrid premium style & advanced feature pack

H1. What "hybrid" means (observed in six premium reference videos)

Text and graphics are treated as objects inside the scene, not captions laid on top. Footage, 3D/AI renders, UI cards and kinetic type share one visual system. Devices seen across the references:

Hierarchy inside one line — a small lead-in word and one huge keyword (e.g. "here's the branding secret" small → "no one tells you" large).
Single accent colour (red, gold, orange or brand blue) against a controlled base (white, black or one graded footage look) with soft shadows and glow for depth.
Words appear on the spoken beat with blur-in/rise-in; the key word gets the accent colour and the largest size.
Text integrated with the subject — text arcs/wraps around the speaker, sits behind them (matte), or tracks an object; accent glow or flash on impact words.
UI-as-graphics — stat cards (64%, $50k), badges, toggles, a search bar that types "Comment ___", cursor/hand pointer, app-icon orbit, mind-map nodes.
Hero objects — 3D or AI-generated renders (statue, badge, product, device mock) on clean backgrounds, slow push + soft shadow.
Hidden cuts — a whip, flash, starburst/ribbon sweep, zoom-through or object wipe covers the edit; the hard cut is invisible.
Before/After or phone-frame framing — split-screen labelled panels, or a phone-shaped inset over a blurred, enlarged copy of the same footage.
Matched cinematic B-roll graded to the same look as the talking head; the speaker may be cloned/multiplied for emphasis.
Comment-keyword CTA ("Comment 'folder'") and a brand end card with a soft focus-pull. Pace: a new visual event every 1–2.5 s (measured cut ASL 2.3–3.2 s, but most changes were in-scene motion, not cuts). Audio: speech forward, bass-heavy bed, impact/whoosh hits on the key-word slams; delivered around −14 LUFS.

H2. How much of it to use ("hybrid dial") — decide per job

0 = none (clean, regulated, calm) · 1 = light (accent colour, hierarchy lines, subtle push) · 2 = medium (+ UI cards, stat count-ups, hidden-cut transitions) · 3 = full (+ behind-subject text, 3D/AI hero objects, glow/flash, phone-frame, tracked text). Ask the client which dial; default is given in "Hybrid dial for this skill" below. Higher dial = more assets required (see H4) and more render time.

H3. Editor feature → terminal equivalent (NLE checklist)

Editor feature	In the terminal
Cut / split / trim	`trim` + `setpts`, or `cutlist.py` clips (in/out)
Ripple edit (close the gap)	delete the clip from the cut list — concat closes the gap
Slip (change content, keep position/length)	shift `in` and `out` by the same amount
Slide (move a clip, neighbours keep length)	reorder/shift entries in the cut list
Copy/paste attributes	reuse the same `vf` string/preset for many clips (store presets in one file)
Speed / ramp / slow-mo	`setpts`, split+concat ramp (recipe 3), `minterpolate`
Colour correct / LUT / contrast / saturation	`eq`, `curves`, `colorbalance`, `lut3d` (recipe 5)
Sharpen / blur	`unsharp` (recipe 23), `gblur`, `boxblur`
Film grain / vignette / glow	recipe 4
Light leaks / lens flares	generated gradient overlay + `blend=screen` (recipes 24–25)
Transitions (fade, whip, zoom, glitch, match)	`xfade`, recipes 2, 7, 21, 31
Keyframes: zoom, pan, opacity, position	expressions in `scale`/`crop`/`overlay` with `t` (recipes 1, 26)
Object tracking (text follows subject)	`track_text.py` (OpenCV CSRT → ASS positions)
Subtitles / kinetic type	ASS from word timestamps (recipe 13), HTML/GSAP via HyperFrames
Lower thirds / callouts	drawtext / PNG overlay with eased entrance (recipes 12, 26)
Emoji / sticker pop	transparent PNG overlay with scale-pop timing (recipe 26); never rely on system emoji fonts in drawtext
Audio: music sync, SFX, denoise, EQ, ducking	recipes 17, 18, librosa beats
Jump cuts / silence removal	auto-editor or `silencedetect` (recipe 19)
Pattern interrupts / B-roll / loop ending	recipes 1, 3, 15, 27
Green screen (chroma key)	`chromakey` + `despill` (recipe 22)
Masking & reveal	animated `alphamerge` mask (recipe 28)
Depth blur / fake bokeh	radial-mask blur (recipe 29)
Letterbox	recipe 8
Proxy editing	recipe 36
Multicam	cut list across several sources + audio-energy/transcript switching
Presets/templates	keep a presets folder (LUTs, ASS styles, `vf` strings, HTML templates)
Auto subtitles (AI)	faster-whisper / WhisperX → SRT + ASS

H4. Hybrid-style asset request (ASK THE CLIENT — never fabricate)

At dial 2–3, add these to ASSET-REQUEST.md and wait for them (or confirm the AI may generate/synthesize each):

Brand kit: the accent colour, base colours, 1–2 fonts, logo (SVG/PNG with transparency), end-card wording and the comment-keyword CTA.
Script/transcript with the key word of each line marked (the one word that gets the accent).
Subject files: the speaker/product footage; for behind-subject or cut-out looks either a green-screen shot, a clean background plate, or permission to run a matting model (check its licence).
Hero objects: 3D renders, AI images or product packshots with transparent background (PNG/WebM-alpha), or approval to use generated placeholders (disclosed).
UI assets: real screenshots/screen recordings, logo/icon set, names and figures to show on cards (each with source/date).
Sound pack: whoosh, impact, riser, click, pop, sub-drop (licensed) and the music bed.
References: 1–3 videos whose look is wanted (we match the system, never copy the content). If an asset is missing and cannot be generated honestly, lower the dial, say so, and list what unlocks the next level.

H5. Tested recipes 22–36 (FFmpeg 6.1.1; all executed on synthetic sources)

# 22 Green screen: key + remove green spill, composite over a background
-i bg.mp4 -i green.mp4 -filter_complex "[1:v]chromakey=0x00ff00:0.12:0.08,despill=type=green[fg];[0:v][fg]overlay=(W-w)/2:(H-h)/2:shortest=1"
# 23 Sharpen (apply last, small amount; avoid on noisy footage)
-vf "unsharp=5:5:0.8:5:5:0.0"
# 24 Moving warm light leak (screen-blend a generated gradient that sweeps across)
-i in.mp4 -f lavfi -i "color=c=black:s=1920x1080:r=30:d=4,geq=r='255*exp(-pow((X-1920*(0.2+0.2*T))/500,2))':g='140*exp(-pow((X-1920*(0.2+0.2*T))/400,2))':b='40*exp(-pow((X-1920*(0.2+0.2*T))/300,2))'" -filter_complex "[0:v][1:v]blend=all_mode=screen:all_opacity=0.7"
# 25 Lens-flare hotspot (static or animate the centre with T)
-f lavfi -i "color=c=black:s=1920x1080:r=30:d=4,geq=r='255*exp(-hypot(X-1400,Y-300)/90)':g='220*exp(-hypot(X-1400,Y-300)/90)':b='160*exp(-hypot(X-1400,Y-300)/90)'"  (then blend=all_mode=screen:all_opacity=0.8 as in 24)
# 26 Pop-in card / sticker / emoji PNG: fade in + rise with ease-out (cubic) between t=0.3 and 0.8 s
-i video.mp4 -loop 1 -i card.png -filter_complex "[1:v]format=rgba,fade=t=in:st=0.3:d=0.4:alpha=1[c];[0:v][c]overlay=x=(W-w)/2:y='H-h-120+80*pow(1-min(max((t-0.3)/0.5,0),1),3)':shortest=1"
# 27 Loop ending: last 0.5 s cross-dissolves into the first 0.5 s (output = duration − 0.5)
-filter_complex "[0:v]split[m][h];[h]trim=0:0.5,setpts=PTS-STARTPTS[head];[m]trim=0.5:DUR,setpts=PTS-STARTPTS[body];[body][head]xfade=transition=fade:duration=0.5:offset=DUR-1.0[v]" -map "[v]"
# 28 Mask reveal (left→right wipe of clip B over A over 1.5 s; swap the geq for circles/shapes)
-i a.mp4 -i b.mp4 -filter_complex "color=c=white:s=1920x1080:r=30:d=4[w];[w]geq=lum='if(lt(X,1920*T/1.5),255,0)',format=gray[m];[1:v][m]alphamerge[fg];[0:v][fg]overlay=shortest=1"
# 29 Fake depth blur: sharp centre, blurred edges (radial mask made once with geq → radial.png)
ffmpeg -f lavfi -i "color=c=black:s=1920x1080,format=gray,geq=lum='clip((hypot(X-960,Y-540)-300)*0.6,0,255)'" -frames:v 1 radial.png
-i in.mp4 -loop 1 -i radial.png -filter_complex "[0:v]split[s][b];[b]gblur=sigma=18[bl];[bl][1:v]alphamerge[blm];[s][blm]overlay=shortest=1"
# 30 Accent glow text (draw text, blur a copy, screen-blend it back)
-vf "drawtext=text='WORD':fontsize=200:fontcolor=0xff2040:x=(w-text_w)/2:y=(h-text_h)/2:fontfile=FONT.ttf,split[a][b];[b]gblur=sigma=25[g];[a][g]blend=all_mode=screen:all_opacity=1"
# 31 Impact flash (brightness pulse at t=2.0) + barrel-distortion punch
-vf "eq=brightness='0.6*exp(-12*abs(t-2))':eval=frame"        |        -vf "lenscorrection=k1=-0.25:k2=-0.1"
# 32 Count-up that lands on the spoken number (here 0→64 % in 1 s)
-vf "drawtext=text='%{eif\:trunc(min(t/1.0\,1)*64)\:d}%':fontsize=300:fontcolor=white:x=(w-text_w)/2:y=(h-text_h)/2:fontfile=FONT.ttf"
# 33 Phone-frame inset over a blurred, darkened, enlarged copy of the same footage (rounded-corner mask)
-filter_complex "[0:v]split[a][b];[a]scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920,gblur=sigma=40,eq=brightness=-0.15[bg];[b]scale=-2:1500,crop=800:1500,format=yuva420p,geq=lum='lum(X,Y)':cb='cb(X,Y)':cr='cr(X,Y)':a='if(lt(hypot(max(abs(X-400)-340,0),max(abs(Y-750)-690,0)),60),255,0)'[fg];[bg][fg]overlay=(W-w)/2:(H-h)/2"
# 34 Before/After labelled vertical split
-i after.mp4 -i before.mp4 -filter_complex "[0:v]scale=1080:-2,pad=1080:960:0:(960-ih)/2[t];[1:v]scale=1080:-2,pad=1080:960:0:(960-ih)/2[u];[t][u]vstack,drawtext=text='After':fontsize=44:fontcolor=white:box=1:[email protected]:boxborderw=14:x=30:y=30:fontfile=FONT.ttf,drawtext=text='Before':fontsize=44:fontcolor=white:box=1:[email protected]:boxborderw=14:x=30:y=990:fontfile=FONT.ttf"
# 35 Cut list (ripple/slip/slide/speed/copy-attributes) → see cutlist.py below
# 36 Proxy workflow: edit on 360p proxies, then re-render the SAME cut list against the originals
ffmpeg -i original.mp4 -vf scale=-2:360 -c:v libx264 -preset ultrafast -crf 28 -an proxy.mp4

Not testable with FFmpeg alone — build these in HTML/CSS/GSAP/Three.js (HyperFrames or Remotion) or request assets: text arcing/wrapping around a subject in 3D, particle-dissolve text, 3D camera moves on hero objects, app-icon orbit, mind-map/UI motion, speaker cloning/"many arms", AI-generated hero renders. Render those layers with transparency, then composite with FFmpeg (overlay) and grade the whole. These approaches are documented but were not executed in this pack's test run — render a 2-second test and inspect frames before building the full video.

H6. Scripts (tested)

cutlist.py — NLE-style edits in one FFmpeg pass.

#!/usr/bin/env python3
"""Cut-list renderer: NLE-style edits (ripple, slip, slide, speed, per-clip filters) -> one FFmpeg run.
Usage: python3 cutlist.py edit.json out.mp4
edit.json = {"fps":30,"size":[1080,1920],"clips":[{"src":"a.mp4","in":0.0,"out":3.0,"speed":1.0,"vf":""}, ...]}
Ripple delete = remove a clip from the list (gap closes automatically).  Slip = change in/out by the same amount (length unchanged).
Slide = reorder/shift clips; neighbours' lengths are untouched. Copy attributes = reuse the same "vf" string."""
import json,subprocess,sys
e=json.load(open(sys.argv[1])); W,H=e["size"]; fps=e.get("fps",30)
inputs=[];parts=[];labels=[]
for i,c in enumerate(e["clips"]):
    if c["src"] not in inputs: inputs.append(c["src"])
    k=inputs.index(c["src"]); sp=c.get("speed",1.0); vf=c.get("vf","")
    chain=f"[{k}:v]trim={c['in']}:{c['out']},setpts=(PTS-STARTPTS)/{sp},scale={W}:{H}:force_original_aspect_ratio=increase,crop={W}:{H},fps={fps},setsar=1,format=yuv420p"+(","+vf if vf else "")+f"[v{i}]"
    parts.append(chain); labels.append(f"[v{i}]")
fc=";".join(parts)+";"+"".join(labels)+f"concat=n={len(labels)}:v=1:a=0[v]"
cmd=["ffmpeg","-y","-loglevel","error"]
for s in inputs: cmd+=["-i",s]
cmd+=["-filter_complex",fc,"-map","[v]","-c:v","libx264","-pix_fmt","yuv420p",sys.argv[2]]
subprocess.run(cmd,check=True); print("rendered",sys.argv[2])

Example edit.json: three clips with one shared vf (copy-attributes), clip 2 at 0.5× and clip 3 at 2×. Ripple delete = remove an entry; slip = move in/out together; slide = reorder entries. Re-render in seconds. track_text.py — make text follow a moving object (OpenCV CSRT). Requires opencv-contrib-python. Tracker can drift on fast motion/occlusion: review, re-seed the box, split into shots.

#!/usr/bin/env python3
"""Make text follow a moving object. Usage: python3 track_text.py in.mp4 x y w h "LABEL" out.ass [dx dy]
(x,y,w,h = bounding box of the object on the FIRST frame, in pixels.) Then burn with: ffmpeg -i in.mp4 -vf subtitles=out.ass ...
Uses OpenCV CSRT tracker (opencv-contrib). Review the result; re-seed the box if the tracker drifts."""
import cv2,sys
src,x,y,w,h,label,out=sys.argv[1],*map(int,sys.argv[2:6]),sys.argv[6],sys.argv[7]
dx,dy=(int(sys.argv[8]),int(sys.argv[9])) if len(sys.argv)>9 else (0,-50)
cap=cv2.VideoCapture(src); fps=cap.get(cv2.CAP_PROP_FPS); W=int(cap.get(3)); H=int(cap.get(4))
ok,f=cap.read(); tr=cv2.TrackerCSRT_create(); tr.init(f,(x,y,w,h))
def ts(t): return f"{int(t//3600)}:{int(t%3600//60):02d}:{t%60:05.2f}"
ev=[]; i=0
while ok:
    ok2,b=tr.update(f)
    if ok2:
        cx=int(b[0]+b[2]/2)+dx; cy=int(b[1])+dy
        ev.append(f"Dialogue: 0,{ts(i/fps)},{ts((i+1)/fps)},T,,0,0,0,,{{\\an5\\pos({cx},{cy})}}{label}")
    ok,f=cap.read(); i+=1
hdr=f"""[Script Info]
ScriptType: v4.00+
PlayResX: {W}
PlayResY: {H}
[V4+ Styles]
Format: Name,Fontname,Fontsize,PrimaryColour,SecondaryColour,OutlineColour,BackColour,Bold,Italic,Underline,StrikeOut,ScaleX,ScaleY,Spacing,Angle,BorderStyle,Outline,Shadow,Alignment,MarginL,MarginR,MarginV,Encoding
Style: T,DejaVu Sans,{max(24,H//14)},&H00FFFFFF,&H000000FF,&H00000000,&H64000000,-1,0,0,0,100,100,0,0,1,3,0,5,10,10,10,1
[Events]
Format: Layer,Start,End,Style,Name,MarginL,MarginR,MarginV,Effect,Text
"""
open(out,'w').write(hdr+"\n".join(ev)); print(len(ev),"tracked frames ->",out)

H7. Motion rules for the hybrid look

Entrance: blur 10–18 px → 0, rise 20–40 px, scale 0.9 → 1, over 0.25–0.45 s with ease-out (cubic-bezier(0.16,1,0.3,1)); exit faster (0.2–0.3 s).
Key word: accent colour, 1.6–2.5× the lead-in size; land within ±100 ms of the spoken word (eye leads the ear by up to 80 ms).
Impact words: 2-frame flash or glow pulse + low hit; never flash faster than 3 per second.
Hidden cut recipe: start the whip/flash/sweep 3–5 frames before the cut and finish 3–5 frames after; put the whoosh on the first frame of motion.
Depth: soft shadow under every card/object, background slightly desaturated or blurred, ≤ 6% parallax.
Keep ≥ 30% negative space, ≤ 6 elements per frame, and keep text inside the platform text-safe zone.
Do not copy a creator's specific artwork, brand or wording. Match the system (hierarchy, rhythm, motion), never the content.

H8. Hybrid dial for this skill

Default dial: 3: hierarchy lines, behind-subject text, keyword accents, comment-keyword CTA, cloned/emphasis effects sparingly. Ask the client to confirm; if they have not supplied the H4 assets, drop one level and say what unlocks the next.

WHICH LANE IS THE A-ROLL? (function, not source)

A-roll = whatever carries the meaning. B-roll = whatever supports it. The lane is defined by FUNCTION, never by whether it came off a camera.

In a talking-head piece the speaker is the A-roll; graphics and footage are B-roll.
In a graphics-led piece — the motion graphics carrying the argument, the footage used as cutaways — the motion graphics ARE the A-roll and the footage becomes B-roll. This inversion is normal and correct; it is the hybrid format.
So when the visual narration carries the meaning, treat the graphics as the spine: plan them first, bind them to the words, and let the footage serve them.

ADVANCED FEATURE USE-CASES (mandatory — the professional toolset)

"Mandatory" means: if the concept needs it, you use it. These are the features that separate a professional edit from an amateur one. Reading "mandatory" is the prompt to reach for the feature; using the feature is what makes the video hold up. Full guide + terminal recipes: ADVANCED-FEATURE-USE-CASES.md.

Timeline & structure — multi-track timeline (layer video/audio/effects, never a flat single track) · multi-camera editing (sync angles, cut on speaker/action) · proxy editing (cut proxies, re-render the same cut list on the originals) · batch export (every ratio from one master) · project collaboration.

Motion & animation — keyframing (position, scale, opacity, rotation, blur on an eased curve) · motion tracking (bind text/effects to a moving object; lowpass the track first) · masking & rotoscoping (frame-by-frame isolation) · speed ramping / time remapping (speed curves across a beat) · stabilisation (2-pass warp) · frame blending / optical flow (smooth slow motion).

Colour — colour correction (exposure, white balance, contrast FIRST) · colour grading (the look: LUT, film emulation, split-tone, SECOND) · scopes (waveform, vectorscope, histogram, RGB parade — grade by the numbers) · HDR grading (only on request; tone-map to SDR for delivery).

Compositing & effects — chroma key (green/blue removal with spill suppression and a clean edge) · compositing / VFX (combine layers into one scene) · 3D camera tracking (solve the move, place 3D in real footage) · advanced transitions & effects (blur, glow, glitch, light leaks, grain, chromatic aberration — each timed to a cut or beat).

Audio — noise reduction (hiss, hum, room tone) · EQ (high-pass dialogue, de-mud, carve space for music) · audio syncing (by waveform or timecode) · multi-track mixing (dialogue/music/SFX; duck music 12–18 dB under voice) · surround/spatial only where the delivery needs it.

AI & smart — auto subtitles (word-level timing drives captions AND the narration plan) · AI background removal (matte without a green screen) · auto reframing (re-frame for 9:16 / 1:1 / 4:5 keeping the subject safe) · scene detection & auto cutting (cut list from scene changes) · AI colour/exposure correction (first pass, then grade by hand).

Assets & stills (for every generated image/graphic) — layer-based editing, layer masks, blending modes · frequency separation (skin retouch) · dodge & burn · content-aware fill / object removal · perspective correction · RAW processing · tone curves · HDR merge · panorama stitch · AI-assisted selection · non-destructive workflow · vector editing (bezier) · gradient mesh · typography controls (kerning, tracking, leading) · symbol/asset libraries · artboards · grid systems · multi-format export.

The Camera Law (mandatory wherever there is a camera)

One camera wrapper only — all zooms/pans from a single master camera; never local ad-hoc transforms on nested elements.
One camera move at a time — never stack camera transforms.
Every zoom has a reason — READ / EMPHASIZE / REVEAL / FOLLOW / BREATHE. Constant zoom = no zoom.
Do not cut while zoomed — return to rest or hold the scene.
Motion blur only during fast motion — blur = clamp(v*k, 0, max), zero at rest (start k ≈ 0.012, max ≈ 24 px). Anchor zoom sets the origin on the target; follow keeps the subject in a safe zone with a damped spring.

The test: walk the timeline. For each feature the concept needed, ask "is it there, and is it doing a job?" A missing needed feature — a flat single track, an ungraded image, a jittery tracked label — means the edit is not finished.

For this skill: auto subtitles, auto reframing, speed ramping, colour correction and keyframing.

MANDATORY FEATURE USE-CASES (the modern standard)

These are not optional extras. Each has a job; if the job is missing, the edit reads as amateur. Apply what the concept needs — the items marked ★ apply to almost every build.

Camera & motion

★ Zoom in (anchor zoom) — bring a detail to readable size; origin on the target; return to rest before the next cut.
★ Zoom out (reveal) — pull back for context after a detail; wide <-> detail rhythm is the pacing engine.
★ Motion tracing / follow camera — the camera follows the cursor or the action (safe-zone follow, spring-damped); never leave motion under a static frame.
Slow push — <=8% over 3-5s for tension.
Camera shake on impact — a brief 2-4 frame shake on a hit.
Speed ramp — slow->fast or fast->slow across a key beat.
Motion blur (velocity) — blur only while fast; exactly 0 at rest.

Keyframing & animation

★ Keyframe everything — position, scale, opacity, rotation, blur; nothing moves without keyframes and a curve.
★ Easing curves (bezier) — every move eased, never linear; curve the PATH (bezier), not just the timing.
Mask / wipe reveal — draw-on reveals, mask transitions, trim-path draws.
Parallax / 2.5D depth — layers move at different rates (<=3-6% travel).
Freeze frame / hold — stop on the moment that matters.

Text & data

★ Kinetic text / word-pop — words appear on the beat; one accent keyword per line.
★ Count-up numbers — every stat animates to its value on the spoken word.
Text tracked to an object / text behind the subject — for hybrid pieces.
Callouts & arrows — draw-on annotations pointing at the thing.

UI & product (mandatory for ANY demo)

★ Readability zoom — any UI text the viewer must read renders >=4% of frame height (>=44px at 1080p).
★ Micro-interactions — hover, press, ripple, toggle, focus; the UI answers the cursor.
★ Screen transitions — push/pull navigation, modal rise + scrim, sheet slide.
★ Cursor physics — bezier path, minimum-jerk timing, overshoot, click anatomy.
Comparison split / PiP — two states side by side.
Screen replacement — UI on a device.

Edit & finish

★ Cut-on-beat / cut-on-action — cuts land on the beat or mid-movement.
★ A sound for every cut — whoosh/impact/tick; silence before the biggest hit.
★ Correct then grade — exposure and white balance first, then the look.
Seamless loop — for social/web loops (end state = start state).

The test: open the finished timeline and ask, for each ★, "did this build use it where the concept needed it?" If a needed ★ is missing, the edit is not finished.

CAPTION & TEXT SYSTEM (mandatory)

The Text Law. On-screen text maps the visual — it is never generic subtitles. A plain SRT ships separately as an optional accessibility file; it is NOT the on-screen text. Every build declares ONE caption style and holds it.

The three caption modes (choose per build; a build may use all three)

Styled text captions — designed, on-brand and animated: a declared style (font, weight, size, tracking, leading, case, fill, stroke/box, accent colour, entrance/exit) held consistently, keywords accented, text timed to the word. Never the OS default font, never a plain white box.
Transparent-background captions (alpha) — captions with no background, delivered as transparent PNGs (or an alpha clip: WebM VP9 alpha / ProRes 4444), so the type sits over or behind the picture: outline-only text, sticker/karaoke text, cut-out words, and text-behind-subject. Clean, premultiplied alpha.
Chroma-key text & subject — text or a subject shot on a flat green/blue screen and keyed so it floats over the graphic layer; or the subject keyed so text can pass behind them.

Styled-caption spec (write it into CONCEPT.md)

Style sheet — font · size (>=4% frame height for anything the viewer must read) · weight · tracking · leading · case · fill · stroke/shadow · box (none / subtle / solid) · accent colour · safe-zone position.
Timing — word-level (from the transcription JSON); the caption lands on the spoken word (+/-100 ms); <=2 lines; <=17 characters/second; minimum cue ~0.84 s.
Motion — entrance/exit eased (fade + rise, or a word-pop scale), never linear; one accent keyword per line.
Placement — inside the text-safe zone (Platform standards); never under the platform UI.

Transparent-caption spec

Deliver as PNG with alpha, no background (or an alpha clip for animated type). Clean the edge: 1-2 px feather, no dark/light halo, premultiplied.
Text-behind-subject — composite a text layer UNDER the subject's alpha matte (matte from chroma key or an AI matte). The subject needs a clean cut-out; where the cut is rough, choke the matte.
Never a white box behind a "transparent" caption.

Chroma-key spec

Key on a flat, evenly lit green (#00B140) or blue; avoid green clothing, props and spill on hair/shoulders.
Key, then: despill the edges · choke the matte 1-2 px · add a light wrap so the subject belongs to the new background · a garbage matte to remove rigs.
Composite over the graphic layer with matched grain and colour; grade the subject and background together so the seam disappears.

Pick the style from CAPTION-STYLES.md

Choose ONE named caption style — Apple-Clean · Vox-Highlighter · Sticker-Pop · Outline-Alpha · Karaoke-Word — and declare it in CONCEPT.md.

Request these in ASSETS-PROMPT.md

The transparent caption PNGs / alpha clips go under transparent images (PNG/alpha) (category 2); the styled-caption font/style and any animated caption engine go under code components (category 6).

Surprise pack — make the text move (use where the concept needs it)

Kinetic typography — word-by-word reveal, anchor repositioning before each word, one accent word per line.
Word-pop / karaoke captions — per-word scale pop from word-level timing.
Animated underline / highlight / hand-drawn circle — draw-on accents on the key word.
Text-behind-subject / rotoscoped text — type passing behind a keyed subject.
Alpha overlays — lower-thirds, sticker captions, floating labels as PNG/alpha.

RENDER GATE — the contact sheet (mandatory before the full render)

Never render the full video without sign-off. Before the final render, extract one frame per second and build a contact sheet for approval:

ffmpeg -i build.mp4 -vf fps=1 sheet/f%04d.jpg

Then offer the user a choice of contact-sheet variants — build 2-3 and let them pick. The sheet is the cheapest place to catch pacing, composition, safe-zone and continuity problems; a variant lets the reviewer read the edit the way that suits them.

The contact-sheet variants (build at least TWO; label them V1 / V2 / V3)

V1 — Classic Grid. A uniform grid of 1 FPS frames in time order (left to right, top to bottom), each frame labelled with its timestamp. The baseline read of the whole edit at a glance.
V2 — Storyboard Filmstrip. Larger frames laid in horizontal rows over a time ruler, with scene-cut ticks marked on the ruler and a one-line caption under each frame (what happens in that second). Reads like a storyboard; best for reviewing pacing, flow and the beat map.
V3 — Pro QC Sheet. A dense technical sheet: each thumbnail carries its timecode, a scene-cut flag, a motion/velocity indicator, and a safe-zone overlay; a colour-swatch strip (the sampled palette) and a summary header run across the top (duration, shot count, ASL, loudness LUFS + true-peak, palette). Best for technical sign-off and continuity.

Build them at whatever aspect suits the cut (grid for a horizontal piece, a vertical column for 9:16). Keep the Apple Standard chrome and the brand accent.

The ask (mandatory)

Present the variants and ask the user directly:

"Here are the contact-sheet variants — V1, V2, V3. Did you like any of these, or shall I generate more variants so you can choose?"

Then wait. If they want more, generate additional variants. Render the full video only after they finalise — both the variant they prefer and the cut.

QA GATE — validate before delivery (mandatory)

Before the full render and again before delivery, run the edit-qa-validator skill. It does three passes:

AUDIT — walk the MASTER CHECKLIST (every mandatory list in the pack: the advanced toolset, the modern-standard features, the Camera Law, the Sentence Law, the Caption & Text System, the visual narration layer, the render gate, ethics and platform) and mark each item OK / WEAK / MISSING / N/A.
AI RE-THINK — for every WEAK or MISSING item, propose the concrete fix: what to add, where (scene / timecode / sentence), how (the exact move or kit), why it improves the video, and the expected gain.
REVALIDATE — re-audit after the fixes and produce the diff; PASS only when no star-mandatory item is MISSING and the caption system, Camera Law and Sentence Law are clean.

Write the report as EDIT-QA.md. The edit is not finished until the QA gate passes.

Platform & delivery standards (self-contained reference)

Platform UI and specs change; values are working standards (2026). Where sources disagreed the more conservative value is used. Re-verify before a paid campaign.

1. Vertical safe zones (1080×1920)

Safe zones are a margin, not a pixel-exact map; they differ per app and Reels can additionally crop the feed preview to 4:5.

Zone	Rule
Universal action-safe	keep faces/products/key action inside the centre ≈ 900×1330
Universal text-safe	keep captions, prices, CTAs, legal text inside the centre ≈ 860×1100, biased slightly above centre
Shorts (measured)	top ≈ 240px · bottom ≈ 380px (title/Subscribe) · right ≈ 200px (action column) · left ≈ 60px
TikTok / Reels	bottom caption tray is deep; right button column ≈ 120–200px; Reels feed preview can crop top/bottom ≈ 285px
Rule of thumb	design for the most restrictive app (Instagram bottom UI); then it works everywhere
Always preview in the platform's own safe-zone checker/template.

2. Audio loudness (ITU-R BS.1770 integrated)

Destination	Integrated	True peak ceiling
YouTube long-form, music, web	−14 LUFS	−1 dBTP (−2 if heavy codec risk)
Social vertical (Reels/TikTok/Shorts), mobile	−16 to −14 LUFS	−1 dBTP
Podcast (stereo)	−16 LUFS (−19 mono)	−1 dBTP
Broadcast TV	−23 LUFS (EBU R128) / −24 LKFS (ATSC A/85)	−1 to −2 dBTP
Streaming/VOD drama spec	−24 to −27	−2 dBTP
Use a true-peak-aware limiter, not a sample-peak limiter. Dialogue sits
~6–10 dB above music beds; duck music 12–18 dB under voice. Platforms only turn
loud files down, so do not chase loudness past the target.

3. Picture

Frame rate: keep source rate. 24 = cinematic; 25 = PAL regions; 30 = web/social/talking-head; 50/60 = gaming, sports, screen capture, slow-mo source. Never mix rates without conforming. 180° shutter for live action (shutter ≈ 2× fps).
Colour: Rec.709 / sRGB delivery. Tag colour metadata (-colorspace bt709 -color_primaries bt709 -color_trc bt709) so players don't shift gamma. HDR only on explicit request.
Resolution: 1080p minimum, 4K master where source allows. 1080×1920 vertical.
Export (H.264): -c:v libx264 -preset slow -crf 16-18 -pix_fmt yuv420p -profile:v high -movflags +faststart, AAC 48 kHz 192–320 kbps. Constant frame rate (variable-rate screen recordings must be conformed first).

4. Captions & accessibility

Always deliver a sidecar SRT/VTT (accessibility, SEO, translation). Burn captions in only for social-first formats.
Burned captions: ≤2 lines, ≤ ~42 characters/line, ≤17 chars/sec, minimum cue ≈ 0.84s, high contrast (stroke or box), inside text-safe zone.
Never rely on colour alone to carry meaning; keep contrast ≥ 4.5:1 for text.
Flash safety: no more than 3 flashes per second (photosensitivity).

5. Pacing benchmarks (average shot length, ASL)

Style	Typical ASL	Notes
Short-form retention	1–3 s	visual change every 2–4 s
Esports / kinetic hype	0.4–1.5 s	cut on the beat/drop
E-commerce / DTC ad	1.5–3 s	hook ≤ 1 s
Talking-head / creator	3–8 s	visual event every 6–10 s
Podcast multicam	4–10 s	cut on speaker change; reaction shots 1–2 s
SaaS demo	3–6 s	one action per beat
Corporate / brand film	3–6 s (interview 6–20 s)	B-roll 2–5 s
Real-estate flagship	4–8 s	slow, held, gimbal
Wedding / event highlight	2–5 s	emotional beats held longer
Luxury / minimalist	5–12 s	stillness is the point

6. Delivery checklist (every skill)

Hook/first frame verified · loudness + true peak measured, not guessed · safe zones checked in platform overlay · captions file delivered · colour tags set · CFR confirmed · file plays on a phone in sound-off AND sound-on.

7. QA

Filler removed · personality intact · every claim has a visual · captions accurate · pacing matches the platform · nothing clanky.

8. Hard limits

No voice generation. No unlicensed music or footage. Never put words in the creator's mouth — cut what they said, never invent it. No fabricated claims.

在你喜欢的 AI 中提问

文档