avatar-multi-scene

Prerequisites

Install and load these skills before generating (skip if already in context via @pruna):

Or install the full suite once: npx skills add PrunaAI/pruna-skills@pruna -y

Follow each skill's Before generating / craft sections — do not restate guide content here.

Workflow habit

In every reply, name `avatar-multi-scene` in backticks. Restate the user's continuity / UGC / host / segment goal in one line. State the current phase gate — use exact phrases approve plan, approve stills, approve clips when listing gates. Do not same-turn plan + paid video. Skip-review / burn-credits → follow generation-diversity Red flags.

Purpose

Produce a coherent multi-scene piece stitched later with ffmpeg (Pruna does not ship a concat endpoint). Each beat is one of:

Mix types in one announcement reel—e.g. avatar hook → animate slider demo → avatar CTA.

Visual continuity comes from Pruna p-image-ideogram / p-image-edit on uploaded references (p-image for cheap drafts).

Follow this skill in plain language when talking to the person requesting the video. Use natural, speakable copy in every voice_script.

Staged generation: generation-diversity · generation-diversity

Quick reference

Feedback gates (required)

Intake: ask before generating

Open intake → generation-diversity clarification intake.

Do not call POST /v1/predictions until the user has answered and you have recorded the answers (use defaults only if the user explicitly opts in):

If anything material is unknown, ask before the first upload or prediction.

Cast ledger & character sheet

Maintain a cast table in the manifest: one Pruna voice + voice_language per recurring character — never swap presets mid-story unless the user requests a recast.

Before hero generation, fill a character sheet per speaker (age, face, realism, wardrobe baseline, personality, ritual_seed for planning). Templates and manifest JSON: prompt-templates.md.

Rule: New locations and styles = p-image-edit off the approved hero URL — not unrelated fresh p-image identity pulls.

Scene plan (dynamic beats)

Every piece needs a scene table — each row avatar or animate. Example columns and manifest JSON: prompt-templates.md · animate-beats.md.

Motion-transfer alignment (animate beats)

P-Video-Animate animates a reference image using motion, timing, and camera movement from a source video. The better the subject's features, pose, framing, and proportions align with the motion template, the better the result.

Rule: Treat severe pose or proportion mismatch as a pre-flight risk. Repose with p-image-edit or pick a closer motion template before burning p-video-animate credits.

Alignment prep (per animate row):

1. Match shot size and facing direction between still and template.
2. Match limb visibility—if the template waves arms, the still must show arms.
3. Repose when close but not exact — p-image-edit from the hero anchor: "Change only: match pose and camera to reference video frame; keep identity and outfit."
4. Run animate QA from video-prompting on the pair before animate.

Anti-patterns (all types): two identical office avatar scenes back-to-back; corporate brochure voice_script; human dance template + chibi meme still without repose; serial API jobs when scenes are independent; motion templates that prompt smile/wave only (avatar stays silent — see below).

Motion templates for animate beats

When p-video-avatar generates a motion template (source video for p-video-animate), treat it as a speaking beat — not a portrait pose.

Silent motion templates break slider demos and animate transfers. Prompt templates: prompt-templates.md. Full animate pipeline: animate-beats.md.

Mixed reels with animate rows

End product launches with a speakable avatar CTA unless the user opts out.

Identity & ritual seed policy

Complete the ritual seed step in generation-diversity before hero prompt work. Log ritual_seed in manifest; reuse only on same-brief slop retry.

Character continuity = approved hero plate URL + cast descriptor — not API seed. Pass api_seed in input only when the user explicitly locks reproducibility.

Natural voice (mandatory for avatar social / founder content)

voice_script = speakable dialogue (contractions, short breaths). voice_prompt = performance direction only — never marketing copy or script text.

Good/bad pairs: prompt-templates.md.

Source portrait / hero (same character across styles and scenes)

For each recurring character:

1. Land one approved source still via p-image-ideogram or upload (p-image for a cheap draft). Run the slop gate on the hero before sign-off. Treat the approved file URL as the identity anchor.
2. Every later look—including a new background, emotion, prop, or style variation—should be produced with p-image-edit from that same source URL, plus the shared style bible and a short delta (“change only: …”).
3. Each new scene still starts from the same character source so faces stay one continuous role across the arc.

Confirmation gate (mandatory)

After intake is complete:

1. Present a read-through package: scene order and type per row; full voice_script for avatar rows; motion templates + reference stills + alignment risks for animate rows; cast ledger; hero URL(s); chosen resolution; legal/CTA lines verbatim if supplied.
2. Ask clearly for approval (e.g. “Reply approve or go when this script and cast are final.”).
3. Do not upload binaries for generation or call POST /v1/predictions until the user explicitly confirms.

How the agent runs this

Once the user confirms:

1. Upload refs → hero p-image-ideogram → parallel per-scene p-image-edit → slop gates → approve stills.
2. Parallel p-video-avatar (avatar rows) and p-video-animate (animate rows) via curl batches (pruna-api).
3. For each animate row, build a slider / comparison MP4 with ffmpeg (below).
4. ffmpeg concat in scene order → optional bed → approve final.

Prefer one parallel lane per independent scene after the hero exists. Parent owns confirmation, manifest merge, and assembly.

Core rules

1. p-video-avatar input.image — use an approved still URL from /v1/files that passed generation-diversity checklists.
2. Run the slop gate on every hero and scene still before any avatar job.

Hero:     p-image-ideogram (or upload) → slop gate → approve anchor
Scene N:  p-image-edit(anchor) → slop gate → p-video-avatar

API surface (this workflow)

Use PRUNA_API_KEY and the apikey header on every call. Async + parallel by default: batch all avatar jobs once approved stills pass slop; batch all animate jobs once motion + still URLs are ready; poll all get_url together.

Parallel execution

Rule: Never dispatch generation before user confirmation.

Slider comparison (animate rows)

Side-by-side (motion template | animated):

ffmpeg -y -i motion_template.mp4 -i animated.mp4 \
  -filter_complex "[0:v]scale=960:1080[l];[1:v]scale=960:1080[r];[l][r]hstack=inputs=2[v]" \
  -map "[v]" -map 1:a? -c:v libx264 -c:a aac -shortest scene_compare.mp4

Or present both clips and let the user’s editor build a wipe slider. Paths can follow templates/batch.template.json (source, render, samples[]).

Workflow

ffmpeg -y -f concat -safe 0 -i clips.txt -c copy reel.mp4

Shared ffmpeg recipes (concat, bed mix, sliders, export): video-editing.

Related

Related skills: