The Cast

Behind every workflow is a small cast of specialized agents doing different jobs. You don't talk to most of them directly — you talk to the Manager, and it delegates. But knowing who's who helps when Claude says "the Visual Planner just returned the storyboard" or "delegating to the Image Prompter".

How agents differ from skills

A quick distinction (covered in detail in the Skills chapter):

Agent	Skill
A collaborator with broad responsibility	A small, named capability
Multi-step — decides, coordinates, judges	Recipe-style — does one thing well
You delegate to it with a brief	You invoke it by name
Script Writer, Visual Planner, Image Prompter, etc.	whisper, last30days, tracker-sync, etc.

The Manager

The director. The agent you actually talk to.

• Specializes in: coordinating the whole pipeline — gathering the brief from you, deciding which agent to call when, passing outputs between stages, surfacing decisions you need to make
• You interact with: every interaction. When you say "make me a workflow," the Manager is who hears it
• Returns to you: confirmations, choices to make, summaries of what other agents produced
• Doesn't do itself: writing scripts, building prompts, running generations — it delegates each of those

Script Writer

The agent that writes the words you'll hear in the video.

• Specializes in: producing the spoken dialogue for a video, structured as hook + body + CTA
• Has 5 modes: new script (Mode 5, default), adapt an existing script (Mode 1), translate (Mode 2), follower-growth content (Mode 3), variant rewrite (Mode 4)
• Takes: the brief (product, audience, tone, sales channel), Schwartz awareness level, any compliance constraints
• Returns: a structured script with hook / body / CTA segments tagged
• You see it via: the Manager presenting the script to you for approval

Visual Planner

The agent that decides what the video looks like, scene by scene.

• Specializes in: breaking a script into scenes, deciding camera angles, framing, lighting, B-roll placement, subject pose per scene
• Takes: the full approved script text, visual direction, B-roll density, product info, audience, sales channel
• Returns: a storyboard with two layers — a user summary (1-2 lines per scene, what you approve) and an internal detail version (full scene specs used by the downstream prompt-writing agents)
• You see it via: the Manager presenting the storyboard summary

Image Prompter

The agent that writes the image generation prompts.

• Specializes in: turning storyboard scene specs into image prompts for NanoBanana 2. Applies the Image Prompt Rules (framing, in-frame, POV, etc.)
• Takes: scene specs from the Visual Planner, account / avatar info, product info
• Returns: image prompt files (plain text + JSON), one per reference image group
• You typically don't see it directly — its output ends up wired into the .nbflow

Veo Prompter

The agent that writes the video generation prompts for Veo 3.1.

• Specializes in: speaking-scene Veo prompts (uses the universal talking head template) and B-roll Veo prompts (natural language, no dialogue)
• Takes: storyboard scene details, dialogue (for speaking scenes)
• Returns: video prompt files, one per scene
• You typically don't see it directly

Seedance Prompter

The Veo equivalent for the Seedance 2.0 model (accessed via the Jimeng web UI).

• Specializes in: positive-constraints-only prompts (Seedance doesn't accept negative prompts)
• Used when: you need Seedance instead of Veo for a specific style
• Manual workflow: produces prompt text you paste into the Jimeng UI yourself

PatchWork Importer

The agent that assembles a .nbflow workflow file.

• Specializes in: wiring all the prompt files into a single .nbflow with nodes connected (image gens → Veo3 → Approval gates)
• Takes: all the prompt files from Image / Veo Prompter, scene structure, reference image groups
• Returns: a complete .nbflow file saved to disk
• Has a Patch Mode for editing existing .nbflow files (e.g. during variant work)

Generation Runner

The agent that runs a workflow and produces the actual images and videos.

• Specializes in: executing every gen node in a .nbflow via the G-Labs API, with retries, R2 uploads, and visual QA
• Takes: a .nbflow file path, tunnel URL, API key, optional reference image
• Returns: a -generated.nbflow file with all R2 URLs baked in, plus local copies of every image and clip
• Auto-reruns images that fail visual QA (AI tells)

Researcher (Viral Scanner)

The agent that does content research — trends, hooks, competitor accounts.

• Specializes in: scanning what's working in a niche, analyzing creators, building creative templates
• Uses skills: last30days, creator-scanner, whisper
• Takes: a niche, a brand, a competitor handle, or a specific question
• Returns: a research brief or creative template

Knowledge Updater

The agent that ingests new information into the pipeline's reference library.

• Specializes in: taking raw source material (a video, an article, a transcript, a brand doc) and updating the appropriate agent's knowledge base
• Uses skills: whisper, notebooklm
• Takes: source material + which agents should know about it
• Returns: updates to brand / product / reference docs

Strategist

The agent for content strategy questions above the per-workflow level.

• Specializes in: content brief work, audience research, prehook selection, platform strategy
• You'd ask it: "what hook style should we use for menopause content this month?"
• Returns: strategic recommendations that feed into briefs

Quick reference

When the Manager says "delegating to X", here's what's happening:

Manager says...	What's happening
"Delegating to the Script Writer"	A script is being generated
"Delegating to the Visual Planner"	The storyboard is being built
"Delegating to the Image Prompter"	Image prompts are being written
"Delegating to the Veo Prompter"	Video prompts are being written
"Delegating to the PatchWork Importer"	The .nbflow file is being assembled
"Invoking the Generation Runner"	Generation is starting
"Consulting the Researcher"	Trend or competitor research is happening
"Updating via the Knowledge Updater"	New info is being added to the knowledge base

When you're ready

→ Next: The Tools — the tools you'll actually open and use day-to-day.