SOP: one short talking-head video, topic to published post

Source

This page is generated from references/SOP_SHORT_VIDEO.md in the repository, the same file the engine and the Claude Code skill read.

An end-to-end recipe that chains the workflows for the most common job: a 60–180 s vertical talking-head short with slides, for 小红书 / 抖音 / TikTok / YouTube Shorts / Reels. Expect ~4 hours the first time and ~2 hours once the slides and cover are template swaps (usually by the third video).

$VSTUDIO = repo root. Run everything from the video’s project folder, e.g. <videos>/short_NNN_<topic-slug>/.

Objective

One topic sentence in, and out:

1. a locked script (~60–120 s of speech), linted
2. a pronunciation drill for the narrator (optional, recommended for non-native speakers)
3. 10–15 square slides, ≤3 of them animated
4. a cleaned, edited talking-head video with the slides in a split layout and captions
5. a cover, then a polished master (loudness, optional speed-up, delivery tags) plus one file per platform
6. post copy per platform, and the post live on at least two platforms

Constraints

• One video per session. Don’t batch-produce; each video gets full attention.
• Lock the script before recording. Rewriting during the recording session is how 4 hours become 8.
• Production friction is the enemy. If a phase blows its budget, ship the rough version and improve the next one. Don’t perfect.
• Creative work when fresh. Visual and editing judgement made late at night looks worse the next morning, and you can’t tell at the time. Leave taste-critical steps (hooks, cover, cuts) for a rested session.
• The creator decides taste-critical choices (hook, which words to cut, cover pattern, speed); the tools propose.

Phase 1: topic and premise (15 min)

Write the topic in one sentence, then answer in one or two sentences each:

• What is the ONE thing the viewer should walk away knowing?
• What does the viewer already (wrongly) believe about it?
• What concrete action or mental tool should they keep?
• Why does the creator find it interesting? (Informs the voice, not the script.)

Check: the first answer fits in 15 words. If not, the topic isn’t ready; pick another.

Phase 2: draft the script (45 min)

workflows/preproduction (part A) and workflows/preproduction/references/script_craft.md.

• Structure: hook (3 sentences) → walk-through → the one insight (signposted) → closing pattern A–D.
• Continuous spoken prose; no fragments, no em-dashes.
• Length: ~165–250 words for 60–90 s, ~250–330 for 90–120 s at 165 wpm (中文: count 字 at voice.zh.cpm).
• Write the hook last: you can only hook well once you know what you are hooking into.

Phase 3: read-aloud pass and lint (15 min)

Read the whole script aloud in performance voice. Fix tongue-trippers, disguised bullet lists, AI-tell words, doubled definitions, forward references, filler reassurance. Then:

python3 $VSTUDIO/workflows/preproduction/scripts/lint_script.py SCRIPT.md --platform youtube-shorts --strict   # or douyin, xiaohongshu

Check: lint has no errors and the self-check in script_craft.md §10 is all ticked. Lock SCRIPT.md.

Phase 4: pronunciation drill (15 min, optional)

workflows/preproduction (part B). Mark every word that trips the narrator, cut to ~8 (10 max), write work/drill_words.json, then:

python3 $VSTUDIO/workflows/preproduction/scripts/make_drill.py work/drill_words.json -o work/pronunciation_drill --script SCRIPT.md

Check: every word is in the script; practise the drill 2–3 times before recording.

Phase 5: slides (60 min)

workflows/slides.

1. Plan breakpoints: one slide per 6–10 s of narration (a 90 s short ≈ 10–14 slides). Slide 01 is the hook.
2. cp $VSTUDIO/workflows/slides/templates/slides_vertical.template.html work/slides.html, swap content per slide.
3. Preview in a browser; check every slide fits 1080×1080, nothing overflows, diagrams hold their geometry.
4. python3 $VSTUDIO/workflows/slides/scripts/render_slides.py work/slides.html work/slides/
5. For the 2–3 diagrams where motion is the explanation: python3 $VSTUDIO/workflows/slides/scripts/record_slides.py work/slides.html work/slides/ 06_bars:3.4

Phase 6: record (45 min)

Vertical 9:16, face and shoulders framed so a slide can sit above or beside the speaker. One continuous raw take if possible; restart a sentence rather than stopping the recording (the cleanup step removes restarts). Keep the first second free of anything important (it becomes the cover). Record in a quiet room; loudness is fixed later, noise is not.

Phase 7: clean up the recording (15–30 min)

Remove 气口 (long pauses), fillers (嗯 / 呃 / um / uh), stutter repeats and restarts with the shared cleanup tool:

python -m vstudio.cleanup --help        # see the options; run it on the raw take

Only high-confidence cuts are automatic; the creator confirms everything else (semantic fillers such as 就是 / 那个 / like are often real words). Re-transcribe the cleaned file and check no content word was lost. On the workflows/talkinghead V track the same job is built in (cut pass 1 + strict_pass.py with AUTO / CONFIRM tiers and verify); use that if you edit there.

Phase 8: edit (45 min)

workflows/talkinghead (V track). Slides go in as BROLL entries in the config: mode="split" (slide on top, speaker below; captions follow the face box) or mode="cut" for a full-screen diagram. Pick the style (记笔记 / 精剪 / 混合) and the hooks with the creator. Captions: read every line against the audio; fix recurring term errors in persona subtitles.term_fixes. Check: compose.py config.py preview <t1>,<t2>,<t3> stills at a hook, a split slide and the ending before the full render.

Phase 9: cover (30 min)

workflows/cover. Pattern B (face over four slides) builds recognition for a new channel; pattern A (collage) shows content density once the channel is known; the talkinghead cover.py gives the 记笔记-style cover. The headline is the script’s hook sentence. Output work/cover.png (1080×1920) or the platform sizes.

Phase 10: polish and export (5–10 min)

workflows/polish, then vstudio.export for the per-platform files:

python3 $VSTUDIO/workflows/polish/scripts/polish.py edit.mp4 -o master.mp4 --cover work/cover.png --speed 1.2 --check
python -m vstudio.export master.mp4 --platforms douyin,xiaohongshu:vertical,youtube-shorts --out exports/ --cover work/cover.png [--cues cues.json]

Propose 1.2× for shorts and ask before applying (above ~1.3× English speech starts to sound artificial; persona speed.cjk_max_intelligible for Chinese). For several platforms, export a caption-free master and pass --cues so each platform gets captions in its own caption box. Check: frame 0 is the cover; loudness within ±1 LU of target, true peak ≤ -1 dBTP; resolution and fps equal the source; exports/manifest.json has no unread warnings.

Phase 11: publish (30 min)

Post copy: workflows/talkinghead/scripts/caption.py config.py --title "..." --platform <name> (title length, chapter timeline in final-video time, tags), method in workflows/talkinghead/references/publish_caption.md; from Python, vstudio.publish.post_body / check_title; or pass --post post.json ({title, hook, body, chapters, links, tags}) to python -m vstudio.export to get a <target>.post.md per platform next to each file.

Platform	Cover	Caption	Hashtags
YouTube Shorts	upload the cover as the custom thumbnail	hook line + one supporting sentence	0–2
TikTok / 抖音	“select cover” → upload the cover	hook line	3–5, niche-relevant
Instagram Reels	“cover” → add from camera roll	hook line	3–5
小红书	custom cover	hook title + body + chapter timeline	per persona publish.tags (most important first)

Check: after upload, the published post really shows the custom cover (some platforms swap back to a frame), and the Short is recognised as a Short (9:16, ≤ 180 s).

Phase 12: next idea (10 min)

The minutes right after publishing carry the most momentum. Write three candidate topics, pick the strongest, write its one-sentence premise, and book the Phase 1–2 session.

Time budget

Phase	Budget
1 Premise	15 min
2 Draft	45 min
3 Read-aloud + lint	15 min
4 Drill	15 min
5 Slides	60 min
6 Record	45 min
7 Clean up	15–30 min
8 Edit	45 min
9 Cover	30 min
10 Polish + export	5–10 min
11 Publish	30 min
12 Next idea	10 min
Total	~4.5–5 h first time, ~2 h when fluent

Failure modes

Symptom	Likely cause	Fix
Final video feels boring	pace too slow, no motion, no music	music bed around -28 dB under the voice (audio.mix_bed), 1–2 animated slides, consider 1.1–1.2×
Quiet on a phone	loudness step skipped or wrong target	re-run polish; check the printed integrated LUFS
Cover looks off-brand	wrong template or accent	keep persona cover.accent = slides.accent; re-render
Captions out of sync	speed applied after captions were burned in	export a caption-free master, polish, then let vstudio.export --cues burn captions
A word disappeared	an aggressive filler cut	re-run the cleanup verify step; take that cut out
Platform shows a random frame as cover	custom cover not uploaded at publish time	edit the post, upload the cover again
Short not in the Shorts feed	over 180 s or not 9:16	trim to ≤ 180 s, export 9:16

Output contract

short_NNN_<topic-slug>/
├── SCRIPT.md                        # locked, linted
├── work/
│   ├── pronunciation_drill.{m4a,md} # optional
│   ├── slides/slide_NN_*.png        # + ≤3 slide_NN_*.mp4
│   └── cover.png                    # or cover.<platform>-<orientation>.png
├── edit.mp4                         # edited talking head (from talkinghead or another editor)
├── master.mp4                       # polished
└── exports/                         # one file + cover per platform, manifest.json

Published on at least two platforms within a day of the polish step.

If you record or edit in Descript (or CapCut / 剪映)

The same phases apply; only 6–8 move into the editor. Use a 9:16 project with a split layout (e.g. Descript’s chapter-title split) and drop the square slide PNGs / MP4s into the slide region at their timestamps. Its built-in filler-word removal replaces Phase 7 (still review every removal). Export 1080×1920, source frame rate, high quality, AAC audio, and without burned captions if you will speed up or export to several platforms (otherwise captions end up doubled or out of sync). Then continue at Phase 9 with polish.py on that export: it replaces the dead first second (often a black title card) with the cover and brings the usually quiet export (around -20 LUFS) to target.