Skip to content

Pipeline Overview

This is the textbook chapter — read The Basics first if you haven't.

A workflow goes from a brief in your head to a finished video in six stages. You approve at two points (after the script, after the storyboard). After that, everything runs automatically — you come back at the end to review the candidates and pick the best ones.

flowchart LR
    S1[1. Brief] --> S2[2. Script]
    S2 -.your approval.-> S3[3. Storyboard]
    S3 -.your approval.-> S4[4. Workflow file]
    S4 --> S5[5. Generation]
    S5 --> S6[6. Review]

Stage 1 — Brief

You tell Claude what you want to make. It asks for the things it needs:

  • Product — what brand, what SKU
  • Audience — who's the viewer (age, gender, life stage, pain points)
  • Tone — warm / authoritative / urgent / conversational
  • Platform — TikTok / Reels / Facebook (defaults to TikTok)
  • Sales channel — TikTok Shop / Amazon (via keyword) / Meta Shop. This decides the call-to-action and whether you need the AI Label Trick.
  • B-roll density — how many cutaway clips. High / Medium (default) / Low / None.
  • Visual direction — what the avatar looks like, what the setting is, lighting style, camera feel

You don't have to answer everything. If you skip a field, Claude proposes a sensible default and asks if it's fine. The whole brief takes 2-5 minutes if you're prepared.

Once the brief is complete, Claude summarizes it back to you for confirmation. That's your last cheap chance to catch a misunderstanding before scripting starts.

What's happening behind the scenes

The Manager (the director agent) loads relevant brand and product docs from the reference library to ground the rest of the work. It also loads the avatar reference sheet for the account you're targeting, and any creative templates the Researcher has on file. None of this requires your input — it just happens.

Stage 2 — Script

After you confirm the brief, ask Claude to write the script. You don't have to specify a script "mode" or framework — Claude picks based on the brief.

A minute or two later, you'll see the script come back. Typically it's structured as:

  • Hook — the first 2-3 seconds, designed to stop the scroll
  • Body — the middle, where the story or argument plays out
  • CTA — the call to action at the end, matched to your sales channel

Read it. If something's off, just say so — Claude iterates on the script with you until you approve.

What's happening behind the scenes

The Manager hands the brief to the Script Writer agent, which specializes in five script modes (Schwartz-framework hooks, follower-growth content, adaptations, variants, translations). It picks the right mode based on awareness level and applies platform-specific compliance constraints (banned words, approved angles). Output is structured with hook / body / CTA segments tagged for downstream agents.

Stage 3 — Storyboard

Once the script is approved, Claude builds the storyboard — a scene-by-scene breakdown of how the video will look.

You'll see a compact summary like:

Scene 01: opening selfie at kitchen counter, holding empty mug
Scene 02: cutaway to coffee being poured (B-roll)
Scene 03: medium shot at kitchen island, gesturing to camera
Scene 04: tight selfie, leaning forward for emphasis
Scene 05: pulled-back shot, product in frame
...

This is the Video Plan — script + storyboard combined. Approve this, and generation kicks off automatically.

What's happening behind the scenes

The Manager hands the approved script (full text) plus the visual direction to the Visual Planner agent, which decomposes the script into scenes with camera angles, framing, lighting direction, subject pose, and B-roll placement. It returns two artifacts: the user summary (what you see) and an internal detail version with every scene fully specified (used by the prompt-writing agents in the next stage).

Stage 4 — Workflow file

This stage is automatic. You don't have to approve anything during it — once the Video Plan is approved, the system builds the entire workflow file in one pass.

What gets produced:

Artifact What it is
Image prompts One per scene (or one per group of scenes sharing a reference image)
Video prompts One per speaking scene + one per B-roll clip
The .nbflow file The complete workflow file with everything wired together
Reference material checklist What images you need to provide (avatar photo, product photo, etc.)

Once this stage finishes, Claude tells you what you need to upload before generation can run.

What's happening behind the scenes

The Manager runs three delegations in sequence:

  1. Image Prompter agent — writes image prompts for each reference group, applying the Image Prompt Rules
  2. Veo Prompter agent — writes speaking-scene video prompts using the universal talking head template, plus B-roll video prompts in natural language
  3. PatchWork Importer agent — assembles all prompts into a .nbflow file, wires the nodes together (image gens → Veo3 → Approval gates), and saves to projects/{month}/{brand}/{growth|sales}/testing/{workflow}-V0-1.nbflow

The Manager also runs consistency checks across stages — same product details, same aesthetic, same tone everywhere.

Stage 5 — Generation

Once you upload the reference materials, ask Claude to run the workflow. Generation takes anywhere from 5 to 30 minutes depending on the workflow size.

You can walk away — there's nothing for you to do during this stage. Claude will:

  • Generate 4 candidate images per scene
  • Generate 4 candidate video clips per scene
  • Visually check every image for problems (extra fingers, melted faces, etc.)
  • Auto-rerun any image that looks broken — up to 3 attempts
  • Save everything to a results file (called a -generated.nbflow)

When it's done, Claude tells you where the results file is. Open it in PatchWork to see what came out.

What's happening behind the scenes

The Manager invokes the Generation Runner, which:

  • Auto-starts a cloudflared tunnel pointing to G-Labs (the local generation backend on port 8765)
  • Uploads any reference images to Cloudflare R2 for permanent public URLs
  • Wipes stale generation cache from the workflow file
  • Calls NanoBanana 2 for every image gen (with outputCount: 4)
  • Calls Veo 3.1 for every video gen (with outputCount: 4)
  • Runs an automated visual review on each generated image to detect AI tells
  • Auto-reruns any flagged image via regen-nodes.py (up to 3 attempts per node)
  • Exports a -generated.nbflow with every R2 URL baked in

Default concurrency is 3 parallel generations, capped at 5. Failed nodes are logged to errors.json.

Stage 6 — Review

Open the -generated.nbflow file in PatchWork (the web app at https://patchwork-33m.pages.dev/). Each scene shows 4 candidate images and 4 candidate video clips in a gallery. Click the best one. PatchWork remembers your picks.

If no candidate is acceptable for a scene, rerun that node — either with the same prompt (re-roll the seed) or after editing the prompt.

When you're happy with the picks, export the final clips for post-production.

Example session

What a brief-to-review pass looks like from your side:

You: I want to make a sales workflow for Brand XYZ — a magnesium
     supplement targeting women 35-55 with poor sleep. Use Account A
     as the avatar. Warm conversational tone, TikTok Shop.

Claude: Got it. A few brief clarifications:
   - B-roll density — high / medium / low / none?
   - Building from scratch, or do you have a reference video to copy?
   - Any specific creative template from the Researcher to apply?

You: Medium B-roll, from scratch, no template.

Claude: Confirming the brief: sales workflow for Brand XYZ magnesium,
   women 35-55 with sleep issues, Account A, warm conversational tone,
   TikTok Shop (with AI Label Trick), Medium B-roll, from scratch.

   Ready to script?

You: yep, go

Claude: [a minute later]

   Script below. 67 seconds estimated, 8 scenes + 3 B-roll cutaways.

   Hook: "If you're a woman over 40 and your sleep stopped working
   when you hit 45 — this might be the missing piece."

   ...

   [full script follows]

You: the hook is good but the third scene drags. Tighten it.

Claude: [revised script, 62 seconds, scene 3 trimmed]

You: approved, do the storyboard

Claude: [storyboard with 8 scenes + 3 B-roll summary]

You: approved, build it

Claude: [10 minutes later]

   Workflow file saved as XYZS1-V0-1.nbflow.
   Reference materials needed:
   - Avatar photo for Account A (already on file, using it)
   - Product photo: Brand XYZ magnesium bottle, white background

You: [uploads product photo]

You: run it

Claude: [25 minutes later]

   Generation complete. 8 scenes × 4 candidates = 32 stills.
   8 Veo clips × 4 candidates = 32 video clips. 3 B-roll clips run
   manually, will be ready in another 10 min.

   2 images were flagged for AI tells on first pass; both regenerated
   cleanly on attempt 2.

   Results: projects/.../Generations/XYZS1-V0-1-generated-{timestamp}.nbflow

   Open it in PatchWork to review.

The whole pass — brief to ready-to-review — is typically under an hour.