SuperAPI AI Directory

make-short-video

General & other

View repository

Updated 10-31 16:16

About this skill

Make Short Video

Turn one sentence into a finished, source-backed Chinese vertical MP4. Keep the workflow automatic unless the title is ambiguous, facts cannot be verified, a required built-in capability is unavailable, or the user explicitly asks for an approval gate.

Non-negotiable contract

  • Produce 1080x1920, 30 fps, H.264/AAC MP4 unless the user overrides a compatible setting.
  • Target 45-55 seconds and hard-stop below 59 seconds. Respect any shorter user duration.
  • Use the final narration audio as the only timing authority. Derive scenes, captions, transitions, sound effects, progress, and total duration from it.
  • Generate the narration as one continuous file. Never stitch sentence-level TTS clips.
  • Use only Codex built-in web research and image_gen, plus local software and local models. Never call a third-party API or request an API key.
  • Keep generated images free of text, logos, watermarks, QR codes, and UI. Add all text in HTML.
  • Version outputs; never overwrite an earlier render, approved script, source asset, or project manifest.
  • Stop with a resumable checkpoint instead of silently switching to an API-based fallback.

Interpret the request

  1. Detect book mode when the prompt names a book, author, ISBN, or asks for a book introduction/review.
  2. Otherwise use general brief mode.
  3. Parse natural-language overrides for duration, platform, audience, style, voice, music, captions, call to action, and output directory.
  4. Before creating a project or writing brief.json, perform a read-only identity preflight for book mode. Ask one concise question only when a title maps to multiple plausible works or reliable facts are insufficient. Create no checkpoint until the identity is resolved; continue automatically otherwise.
  5. Default to Chinese adult short-video viewers, cinematic editorial storytelling, local Piper Mandarin female voice zh_CN-xiao_ya-medium, a restrained rhythm-free ambience bed, and no hard-sell CTA.

Read workflow.md before starting a new project. Read style.md before writing the script or image prompts. Read project-schema.md before creating project.json. Read qa.md before rendering.

Create the project

Create short-video-runs/<YYYYMMDD-HHMM>-<slug>/ under the active workspace unless the user names another destination. Copy DESIGN.md into the run and specialize it without changing the safe-zone contract.

Create these durable checkpoints:

sources.md
brief.json
angles.json
script.txt
storyboard.json
project.json
assets/images/
audio/
captions/
qa/
renders/

Do not create placeholder media. Preserve every checkpoint that completed successfully so a failed run can resume.

Research and write

Book mode

  • Verify the exact title, author, edition-independent premise, main themes, and any plot facts used in the script.
  • Prefer publishers, author pages, libraries, ISBN catalogs, and reputable interviews or reviews. Record links and claim-level notes in sources.md.
  • Paraphrase. Do not invent quotations, reader reactions, awards, sales, characters, or plot events.

General brief mode

  • Research only when the video makes factual claims. Record the sources used.
  • Treat the user's supplied facts as user-provided context, not independently verified truth.

Generate three internal angles and score each for 3-second hook, factual support, visual potential, naturalness, and fit with the request. Save all three to angles.json; choose the highest score automatically.

Write a natural 180–220 Chinese-character narration for the default duration. Shorten or expand for overridden durations. Avoid hard-sell copy, catalog summaries, arrogant claims, excessive parallelism, and stock AI patterns such as repeated “不是……而是……”. Make the first sentence immediately meaningful.

Build the visual plan

  • Split the narration into 6-8 semantic beats, or 4-6 beats for videos under 35 seconds.
  • Give every beat one purpose, one image concept, at most three overlay keywords, and an anchor phrase that exists verbatim in script.txt.
  • Use Codex built-in image_gen once per scene. Use the first accepted image as the style anchor for subsequent scenes.
  • Prompt every image as native 9:16 with the subject, action, environment, period, light, emotion, palette, header negative space, caption negative space, and consistency constraints.
  • Keep faces, hands, actions, and key objects outside the header and caption bands defined in style.md.
  • Inspect every image. Retry a failed scene once with one targeted correction. If built-in image generation remains unavailable or unusable, stop and preserve the project; do not use the CLI/API fallback.

Generate and align audio

Run bootstrap.ps1 in check mode. Run it with -InstallSystemDependencies -InstallModels only when dependencies are missing and installation is in scope.

Run prepare_audio.ps1 to synthesize all of script.txt in one call with Piper zh_CN-xiao_ya-medium, post-process lightly with FFmpeg, transcribe the final narration, force-align its timing to the confirmed script, generate optional BGM, and create audio/master.wav.

Reject alignment coverage or exact text match below 0.90. Fix pronunciation or regenerate narration instead of guessing timestamps. Do not hide unnatural speech with pitch shifting, time stretching, or heavy compression.

The audio preparation script uses align_transcript.py and generate_bgm.py. Keep the narration dominant and reject alignment coverage below the threshold.

Compose and render

Create project.json only after the final audio and caption timings exist. Map each scene boundary to the first timestamp of its anchor phrase. Keep scenes contiguous and within the narration duration.

Run render_project.ps1 in draft mode. It validates the project, builds the deterministic HyperFrames composition, runs lint plus the current consolidated check pass, renders a draft, generates an ffprobe report, and creates a contact sheet.

Inspect the contact sheet with view_image. Check the opening, every scene's hero frame, the middle, and the ending. Fix real problems once and rerun the draft. Then run render_project.ps1 -Mode final.

Deliver

Return clickable paths to:

  • the newest final MP4;
  • script.txt;
  • project.json;
  • the newest qa/report-final-vN.json and matching final contact sheet.

State the duration, resolution, frame rate, voice, BGM setting, and whether all automated and visual checks passed. Mention any remaining limitation plainly.

Update this skill only on explicit request

Do not self-modify during normal video production. When the user explicitly asks to update the skill from an approved run, add only transferable rules or reusable component notes to library.md, update the relevant reference, validate the skill, and summarize the behavioral change.