Lanshu Create AI Presenter Video

Lanshu Create AI Presenter Video

Produce a verified explainer video from minimal inputs — led by a digital human, or performed in one of nine visual styles. The final narration is the master clock for presenter motion, captions, graphics, cuts, and delivery duration.

Choose a route

Every job takes one of two routes, recorded as creative.route. Infer it when the request is clear (cues below); otherwise ask once, with this table and the style gallery (explainer/STYLES.md; attach or link the image explainer/gallery/nine-styles-kv-cache.jpg). Never fall back to presenter silently: it needs a portrait the user may not have.

• A presenter image or the words 数字人 / presenter / 口播 / 真人出镜 point to presenter; 不要真人 / "no presenter" / "faceless", a named style, or 动画讲解 / 手绘 / 黑板 / 黏土 / 漫画 point to styled.
• A vertical platform (抖音 / 视频号 / 小红书 / 竖屏 / 9:16) is a format request, not a route: styled is 16:9 only. If it comes with no portrait or with styled cues, ask (a 16:9 styled film vs a 9:16 presenter video) instead of switching routes.
• Narration the user already recorded is the locked audio on either route. Never pass it as --voice-sample (that flag is for voice cloning). On styled use story.py story --audio <file> and ask for the matching script text (or transcribe it and have the user check it).
• To re-style an earlier job, keep its script and locked audio (no new voice cost), build a new film from the same story.json in the new style, and re-approve the storyboard. Ask which job when it is not obvious.
• This skill makes explainers. For 宣传片 / 广告 / 发布大片 / footage edits, say so, and offer a product explainer only if the user wants one.
• For styled also choose creative.explainer_style (v1-editorial … v9-clay). Recommend two or three from the topic and audience (one reason each) using explainer/STYLES.md; when the user is unsure, build drafts in all nine (no extra paid calls once the audio is locked, or with --dry) and let them pick from a still grid. Settle the style before the script is locked: each style shapes how the narration is written and voiced (explainer/STYLES.md, "按风格写稿"). To compare styles first, use --dry drafts of a first script, then adapt it to the chosen style.
• styled follows styled-explainer.md and reuses this skill's job directory, state machine, approvals, and delivery checks. The styles are designed to carry the explanation on their own; if a user also wants their digital human in a styled film, that file describes the method (time the film to the presenter's locked narration, give the presenter its own space) as an optional addition, not a separate route.

How to ask, when the request leaves the choice open — one message, in the user's language:

1. The two routes, one line each: what the video looks like, what the user must provide, what is paid, and the format (presenter: any ratio, 9:16 by default; styled: 16:9 only). Recommend one for this topic and say why (e.g. "no portrait yet, or a concept that needs to be shown → styled"; "personal brand, talking to camera → presenter").
2. For styled, assume the user does not know the styles yet: show the whole menu — the gallery image plus the nine one-line entries of "风格菜单" in explainer/STYLES.md — and mark the two or three that fit this topic and audience, one reason each. Offer "build drafts in all nine and pick from a still grid" as the alternative. Do the same, without the route table, when the user only asks which styles there are.
3. Proceed with the recommendation if the user agrees or does not mind; record the choice in job.json.

Required outcome

• Start from a topic or script and, for the presenter route, one authorized image containing one clear adult presenter.
• Deliver a fully decoded master video, a smaller share copy, captions, production records, and separate machine and visual QA notes.
• Keep the workflow portable across providers, models, aspect ratios, languages, and durations.

Operating rules

• Confirm image rights, adult status, remote-upload approval, and voice-cloning authorization before the relevant remote action.
• Treat remote generation as potentially billable. Before the first paid call, state the uploaded assets, requested seconds or units, known cost, pilot size, retry ceiling, and expected output. Reuse an approval already given for that exact plan.
• Never infer a real voice from an image. Use an authorized sample or record a selected stock voice.
• Lock the complete narration before presenter generation, caption timing, or final scene boundaries.
• Prefer one presenter image, one voice identity, one visual treatment, and one continuous presenter source.
• Mute video sources in the final composition. Route only the approved external narration and intentional mix tracks.
• Preserve provider request bodies and task IDs without credentials or expiring URLs. Poll interrupted work before considering resubmission.
• Use numeric checks for technical faults and normal-speed visual review for identity, mouth timing, blinking, gestures, hands, lighting, and continuity.
• Stop after three rejected paid candidates, or before a change that materially affects cost, privacy, voice, appearance, or provider.
• Do not claim completion until the final files fully decode, the contact sheet or full playback has been reviewed, and check_state.py reports verified.

Start or resume a job

For a new job, read generation.md. Set SKILL_DIR to the directory that contains this file; its location depends on the harness. Then run:

SKILL_DIR=/path/to/lanshu-create-ai-presenter-video

python3 "$SKILL_DIR/scripts/init_job.py" \
  --job-dir ~/Videos/my-presenter-video \
  --presenter-image ~/Pictures/presenter.png \
  --topic "用一分钟讲清楚上下文工程"

For the styled route add --route styled --explainer-style v3-notebook (no presenter image needed).

Use --script for an existing script file. Optional flags include --voice-sample, --supporting-media, --duration, --aspect, --width, --height, --fps, --watermark, and --cta. The initializer copies every input into the job and records job-relative paths, so the job directory is self-contained; keep later artifact paths job-relative too.

Inspect the actual source image and listen to any voice sample. Record the manual review and approvals in job.json, then run:

python3 "$SKILL_DIR/scripts/preflight.py" ~/Videos/my-presenter-video/job.json

For an existing job, read job.json, current artifacts, task IDs, and QA reports, then run check_state.py (below) to find the earliest unfinished state. Resume there without regenerating accepted work.