Video Production

What this skill is for

Turn raw footage into a finished short vertical video that someone would actually post: a hook, a clear story, readable burned-in subtitles, intentional packaging, clean audio, and a verified delivery package. It also protects an already-approved cut when the user asks for changes.

The goal is not to copy one successful video's shots or filters. The goal is to reuse the conditions that made it good:

a written production constitution before editing;
timestamped material evidence, not vibes;
an executable shot script;
a deliberate visual system (layout, typography, safe areas);
independent review at the material / script / final gates;
objective technical QC;
versioned, checksummed delivery.

When to use

The user supplies clips and asks for a vlog / short vertical video / campus or travel piece.
The user asks for burned-in subtitles, titles, fact cards, an end card, or a cover.
The user asks to grade flat/Log footage or match shots.
The user asks to revise an approved video: change subtitles, music, opening, packaging, a few shots, or replace footage.
The user asks why a previous cut was good/bad and wants a reusable process.

When not to use

The user only wants a written concept, script, or shot list — no edit.
Long-form film/TV finishing, color-grading a feature, or complex VFX/3D.
Pure audio/podcast editing.
The request is a tiny text-only answer that needs no media inspection.

Step 0 — Classify the job (do this first)

Class	Signals	Process
New production	new topic, new clips, new deliverable	full workflow
Local patch	only subtitles, title, a few overlays, one B-roll, volume, music fade	minimal diff + A/B + re-review
Full redo	replace all/most footage, change layout/aspect/visual system, change grade pipeline, rewrite story	treat as new production; do not patch the old timeline

Hard rule: "replace all footage" or "redo it with these new clips" is a new production, even if the user calls it a small fix. The old script's source timecodes, framing, subtitles, and B-roll logic no longer hold.

If it is a revision: before touching anything, freeze the approved version (copy + SHA-256). See references/modification-protection.md.

Step 1 — Intake and production constitution

Capture only what affects delivery:

platform / aspect / target duration / language;
footage paths, Log vs Rec.709, frame rate, audio;
whether there is a talking-head track, and who speaks;
music/licensing (CC0/CC BY/user-provided);
delivery folder/name/overwrite policy;
what must NOT change.

Write a short 00-production-spec.md (or reuse the project's existing one) with:

specs and red lines;
factual baseline (what may appear on screen; what is only a speaker's opinion);
visual grammar: card vs full-bleed, title, subtitles, fact cards, end card, safe areas;
sound plan;
review gates.

Do not skip this because the job looks small. The spec is what lets a reviewer say "this violates the agreed system" instead of arguing taste.

Step 1.5 — Reference study (new production / full redo)

Before choosing the visual system or writing the shot table, study 3–5 high-quality reference vlogs in the same platform/genre/duration. Extract principles, not shots:

first 1–3s hook: image / words / sound;
structure: shot count, average shot length, A-roll/B-roll ratio, fact-card/end-card placement;
packaging: subtitle size/position, title, color, transitions, information density;
sound: voice, BGM, SFX, ambience;
why it works in one sentence;
transferable principle vs surface element.

Use references/reference-study.md for the template. Do not copy shot-for-shot, faces, music, or copyrighted assets. Write a short reference-study.md in the project and use it to make 2–3 creative options before committing. For a local patch, skip this unless the user asks.

Step 2 — Material audit (evidence, not impressions)

For every clip record:

duration / resolution / fps / codec / bit depth / color tags;
content and action by timestamp;
usable in/out points (0.5s precision) and reject reasons;
framing, headroom, logo/background, shake, exposure, white balance;
audio: speech vs ambience; if speech, a timestamped transcript;
Log/HLG or Rec.709; if flat, plan the grade.

Use 1s contact sheets + key frames; use Whisper for a first transcript, then correct proper nouns with context. Do not subtitle B-roll ambience that Whisper hallucinated.

Output: a material report with frames and timecodes, not just prose. A reviewer must be able to re-extract the same frames.

Step 3 — Visual system and shot script

3.1 Choose the A-roll treatment from the source

Card layout for low-res landscape (e.g. 960×720) or when the environment/logo matters: a 4:3/16:9 card on a designed 1080×1920 canvas; avoids 2.67× upscale.
Full-bleed for native vertical or 4K landscape with enough headroom: test the crop before committing.
Hybrid when both face and context matter.

Do not change an already-approved layout unless the user asks or the new source genuinely cannot support it. If you do change it, treat it as a design change and re-review.

3.2 Write a shot table, not a mood board

Every shot needs: source, source in/out, final in/out, duration, picture treatment, subtitle, packaging, transition, sound. Sum the durations; align the voice track by a fixed offset; use final-timecode subtitles.

Include:

a hook in the first 3s;
a visual/information change every 2–4s;
B-roll that covers weak A-roll sections;
fact cards / end card copy with sources and risk labels;
a music/sound plan.

Review the script before editing. A wrong source timecode is cheap to fix here, expensive later.

Step 4 — Build

Render each shot separately; concatenate/xfade in a second pass. Avoid one giant filter graph.
Grade Log footage with a LUT or a tested manual chain; compare face / white wall / brand color across shots.
Generate overlays (title, fact cards, end card, subtitles) as full-canvas RGBA PNGs with Pillow; test them on real frames.
Keep subtitles in the bottom safe area (bottom margin ≥200px; 240px is safer), ≤14 Chinese characters per line, ≤2 lines.
Mix voice first, then BGM under it (12–18dB below voice), fade in/out 1–3s, loudnorm to about −16 LUFS, true peak ≤ −1dBTP.
Keep all scripts, overlays, audio, and intermediate shots versioned.

Useful checks and recipes: references/tech-recipes.md.

Step 5 — Review gates (do not self-approve)

Three independent gates:

Material gate — does the research match the frames/audio? Are usable ranges real? Are ASR errors corrected? Are B/C/D unbilled ambience left alone?
Script gate — can it be edited exactly as written? Do source in/out points exist? Do subtitles match the transcript and fit? Is the total duration exact? Is the hook real?
Final gate — full-film frame sampling + SRT read + audio check + factual check. Technical QC (resolution, fps, duration, loudness, true peak, black frames, decode, burn-in) is separate and cannot replace the visual/subtitle review.

Each finding needs: timecode, symptom, fix, severity (blocker/high/medium/low). A pass requires no blocker/high; medium should be fixed or explicitly accepted; low can ship with a note.

If a review fails, repair and re-review the affected area. Do not deliver on "the checker passes".

Step 6 — Deliver

Final MP4: 1080×1920, 30fps, H.264 High, AAC 48k stereo, faststart.
SRT with the same cues as the burn-in; cover 1080×1920.
交付说明.md: specs, files, sources, structure, subtitles, music/CC BY, factual sources, known limitations, review verdicts.
SHA-256/MD5 manifest; keep the approved version/master read-only.
Confirm QuickTime/mobile playback.

Step 7 — Revisions

Follow references/modification-protection.md. The short version:

freeze the approved version;
list exactly what the user asked to change;
treat everything else as "must preserve";
use the smallest patch that works;
if re-generation is unavoidable, declare the risk and re-run the full review;
A/B compare old vs new at the same semantic beats;
never overwrite the only good version;
if the new version fails, roll back and keep the old one.

Quality bar (defaults; adjust to the brief)

Item	Default
Canvas / fps	1080×1920 / 30fps
Duration	45–60s for a vlog; 15–30s for a hook-led short
Video	H.264 High, yuv420p, CRF ~18, faststart
Audio	AAC 48k stereo, ~192kbps
Loudness	−16 LUFS ±2, true peak ≤ −1dBTP
Black frames	none except intentional design
Subtitles	burned-in + SRT; bottom margin ≥200px; ≤14 chars/line
Music	voice-first; BGM 12–18dB under voice; fade in/out
Delivery	MP4 + SRT + cover + delivery note + checksums

These are defaults, not laws. If the platform or brief requires something else, write it in the spec.

Common failure modes (from a real revision)

Do not change an approved visual system unless asked. An approved cut used a 4:3 card for low-res landscape; a later revision replaced it with a full-bleed 4K crop, and the reviewer complained that the landscape footage had simply been cropped. If the source is landscape and the approved look was a card, keep the card unless the user asks for a redesign; if changing, test headroom/logo/hands and review.
Fact-card backgrounds must be distinct and calm. Do not reuse overlapping source ranges or two near-identical pans behind consecutive cards; slow/freeze the background if the camera moves. A moving, duplicated background reads as "two almost identical shots" and "the camera sweeps too fast".
Every bridge segment needs a function. A 2–3s shot between the last fact card and the end card must carry a title, a voice, a location line, or a narrative beat. A silent, textless banner reads as an unexplained gap; shorten it or give it a role.
A full footage replacement is a new production. Do not patch the old timeline; re-audit material, re-script, re-grade, and re-review. "Small fix" language does not change the scope.
Technical pass is not an aesthetic pass. A checker can prove 1080×1920/0 black frames/decode OK; it cannot prove the cut looks good. For revisions, A/B against the approved cut and get an independent verdict.

Creative latitude

Standardize the verification; leave the creative open:

concept, hook, narrative order;
card vs full-bleed vs hybrid;
shot count, B-roll order, transitions;
grade look (natural / warm / filmic);
music genre, title voice, fact-card visual.

Do not turn one approved cut into a template. The reusable asset is the process and the review discipline, not the exact fonts, shots, or music.

References

references/case-study.md — an anonymized case study: what worked, what regressed, and why.
references/workflow.md — full 7-stage workflow with artifacts and gates.
references/modification-protection.md — versioning, minimal diff, A/B, rollback.
references/quality-checklist.md — pre-flight, technical, visual, subtitle, audio, delivery checklists.
references/reference-study.md — how to study 3–5 reference vlogs and extract principles without copying.
references/tech-recipes.md — ffmpeg/Pillow/Whisper/loudness recipes and known pitfalls.
references/skill-sources.md — Agent Skills format, public examples, and host installation notes (Claude Code, DSH, other hosts).
scripts/check_video.py — reusable technical checker.
scripts/contact_sheet.sh — quick 1s contact sheet.

Quick start

# discover tools
FF=$(command -v ffmpeg || echo tools/ffmpeg)
$FF -hide_banner -i clip.mp4 2>&1 | grep -E 'Duration|Stream'

# 1s contact sheet (5 columns)
$FF -hide_banner -loglevel error -i clip.mp4 \
  -vf "fps=1,scale=320:-1,tile=5x6:padding=4:margin=4" \
  -frames:v 1 contact.jpg

# technical check after export
python3 scripts/check_video.py final.mp4 --width 1080 --height 1920 \
  --min-duration 45 --max-duration 60 --lufs -16 --lufs-tol 2 \
  --max-true-peak -1.0 --max-black 0.5

If the user has an approved version, stop and read references/modification-protection.md before running anything that writes to the deliverable.

YI163/video-production

video-production 是什么？

在你喜欢的 AI 中提问

文档