Pruna generative pipeline (scenario hub)

When NOT to use

• Single-shot generation with minimal intake → pruna-run
• One talking-head clip only → avatar-single-scene

This skill is the menu of proven chains over the atomic tools in tools/image and tools/video. For HTTP details see pruna-api.md.

Marketing and demo sites change layout and messaging often; this repo is the source of truth for what agents should run. When humans find a compelling flow elsewhere, encode it here (extend a recipe, add a row to the idea map, or add a dedicated workflow skill) after checking it against docs.api.pruna.ai—do not send agents to browse live demos as part of execution.

Tone: Explain plans to the user like a producer—clear scene order, what generates first, and what they should hear or see—especially whenever voice_script or narration exists. Avoid stiff jargon in client-facing script drafts unless they ask for it.

Quick reference

Intake: pick a scenario first

Ask the user which recipe fits (or hybrid). Capture answers before any POST /v1/predictions:

Confirmation and execution (all recipes)

0. Generation diversity — ritual seed + diversify axes before the first generation in a session.

1. Draft any scripts, beats, or prompts that the user must approve in natural, human language (spoken wording for VO; clear intent per scene otherwise).

2. Pause and obtain explicit confirmation (“approve”, “go”, “run it”) before the first upload or POST /v1/predictions. If the user revises copy or cast, re-confirm when the change matters for cost or outcome.

3. Staged execution — after plan approval: stills only → user approves → audio prep (TTS / song) when applicable → user approves → video → user approves clips → assembly and final audio (bed). See staged-generation-gate.md and requesting-generation-feedback.

4. Write a runnable generation package—script or phased API steps—that matches the approved plan exactly. Use async parallel fan-out within each phase (see parallel-execution.md); avoid serial scene-by-scene execution when lanes are independent.

5. Execute that package when possible (PRUNA_API_KEY, network). Prefer subagents per scene lane (still pipeline or avatar job) when the host supports parallel agents. If execution is not possible here, deliver the same artifact for local runs.

Deep avatar workflows already spell out cast ledgers, hero reuse, and read-throughs in avatar-multi-scene and avatar-single-scene; defer to those when recipe G or H involves p-video-avatar.

Routing table

Handoff rules (all recipes)

• Use Pruna file URLs only (POST /v1/files → urls.get) in images, image, audio.
• Async + parallel by default for p-video, p-video-avatar, and batch stills—create all jobs in the current phase together, poll all get_url until done (parallel-execution.md). Use subagents for independent scene lanes after confirmation.
• Do not chain the next step until the previous URL is valid and (for portraits) checklist-approved when using recipe G.
• Character continuity: One approved source / hero URL per recurring subject; branch new looks and styles with p-image-edit from that URL (plus style bible), not unrelated fresh p-image identity pulls—unless the user resets the character.
• Dynamic scenes: Multi-scene avatar pieces must vary camera angle, background/setting, lighting, and visual_style_tag per beat—no repetitive office-only stacks unless requested. Run visual-variety-bible.md checklist on launch reels.
• Photoreal stills: Hero via p-image-ideogram (high-quality fast) or p-image (fastest drafts) with documentary/photoreal prompt → slop gate; every p-image-edit still → slop gate before p-video-avatar. Persona bar: realistic-persona-showcase.md.
• Dynamic avatar motion: unique video_prompt per scene — not duplicated medium-CU dolly on every row.
• Identity lock: Record approved hero plate URL at hero generation; branch with p-image-edit for same character. Log ritual_seed (SSoT) for prompt planning — do not pass to API unless user sets api_seed.
• Same voice for the same role: When a character speaks in more than one clip, keep voice (and usually voice_language) identical across those clips.
• Human delivery: voice_script = speakable dialogue; voice_prompt = realistic performance direction (never marketing copy).

Recipes (one-liners)

Full intake, steps, and shine lines for every letter: recipe-catalog.md.

Atomic tool index

Audio layering guide: audio-post-production.md · Scene anchor triple: scene-anchor-triple.md

Product-positioning idea map: recipe-catalog.md More ideas.