Skip to content

Chapter 11 — The Pipeline in Detail

The deep-dive chapter. By now you've worked with workflows, made variants, fanned out, set up accounts, and done research. Before you tackle creating a video from scratch (the most complex task in the pipeline), you need to understand each agent in detail, the visual formats available, and the architecture patterns that make workflows maintainable.

Chapter at a glance

flowchart TD
    A[Agent cast<br/>in detail] --> B[The Prehook]
    A --> C[PiP format]
    A --> D[Storyboarding<br/>logic]
    A --> E[Pipeline<br/>variants]
    A --> F[avoid-ai-writing]

Concepts you need before tackling a video from scratch.

What you'll learn

  • The full agent cast in detail — what each agent specializes in, what they take and return
  • The Script Writer's 5 modes — when each mode applies
  • What a prehook is and when to use one
  • PiP (picture-in-picture) format — mixed-media composition
  • Storyboarding logic — the 8-second Veo clip constraint, word counts per scene, B-roll placement decisions
  • Pipeline variants — AI Self / Avatar content, voice cloning workflow
  • The avoid-ai-writing skill — auditing scripts for AI tells in writing

Before this chapter

Read Chapter 10 — Content Research. The chapters up to here have prepared you with vocabulary, tools, and operational experience. This is the last concept chapter before from-scratch creation.

Sections

  1. Agent cast in detail — Script Writer modes, Visual Planner, Image/Veo Prompter at depth
  2. The Prehook — what it is, when to use, how it fits in the workflow
  3. PiP format — picture-in-picture composition for recipes / explainers
  4. Storyboarding logic — 8-second constraint, word budgets, B-roll placement
  5. Pipeline variants — AI Self / voice cloning configurations
  6. avoid-ai-writing in practice — auditing scripts for AI tells

Architecture details + schema reference moved

.nbflow architecture patterns, the schema reference, and the per-model technical cards used to live in this chapter. They're now in Under the Hood — the reference section for people improving the system, not operators producing videos. If you're trying to use the pipeline, you don't need them.

Try this

Open an .nbflow from a workflow you've shipped and trace every output back to its agent.

Steps:

  1. Open the Assets/{workflow}/ folder for a workflow you've already produced. List the files: video-plan.md, storyboard.md, main-body/scene-NN-image-start.json, main-body/scene-NN-video.txt, etc.
  2. For each file, name the agent that produced it (Script Writer? Visual Planner? Image Prompter? Veo Prompter?). Check against Agent cast in detail
  3. Identify the prehook in the storyboard (if present). What scene number, what's its goal? (The Prehook)
  4. If the workflow uses PiP, identify which scenes use it and why (PiP format)
  5. For one scene's Veo prompt, count the words in the dialogue. Confirm it's under the 8-second word budget (Storyboarding logic)
  6. (Optional) Run the avoid-ai-writing skill on the script. See what it flags (avoid-ai-writing in Practice)

If you hit:

  • A file's producing agent is unclear → re-read Agent cast in detail. Each file belongs to one agent
  • A scene's dialogue exceeds the word budget → it'll get truncated or rushed in the final video. Re-read Storyboarding logic and ask Claude to tighten the line
  • You want to understand .nbflow JSON internals or architecture patterns → that's Under the Hood. Most operators don't need it
  • The workflow uses an unusual pipeline configuration (AI Self, voice cloning) → see Pipeline Variants

When you're ready

→ Next: Chapter 12 — Create a Video From Scratch. Everything you've learned, applied to the full brief-to-delivery flow.