Skip to content

Pipeline Overview

The 6 stages from product brief to delivered videos.

flowchart TB
    S1[1. Brief & Setup]
    S2[2. Script Creation]
    S3[3. Visual Storyboard]
    S4[4. PatchWork File]
    S5[5. Generation]
    S6[6. Delivery & Feedback]

    S1 --> S2
    S2 -.user approves.-> S3
    S3 -.user approves Video Plan.-> S4
    S4 --> S5
    S5 --> S6
    S6 -.feedback loop.-> S2
    S6 -.feedback loop.-> S3
    S6 -.rerun.-> S5

The user approves once after Stage 3 (the Video Plan — script + storyboard together). Stage 4 then runs automatically.

Stage 1 — Brief & Setup

Gather from the user:

  • Product — what brand and SKU. The Manager loads any existing brand and product reference docs to ground the work.
  • Target audience — age, gender, lifestyle, awareness level.
  • Tone / style — aspirational, relatable, edgy, warm, etc.
  • Reference material — existing scripts, viral videos to copy, brand examples.
  • Platform — TikTok / Reels / Facebook. Defaults to TikTok.
  • Sales channel — TikTok Shop, Amazon (via keyword), or Meta Shop. Affects CTA and AI Label Trick requirements.
  • Content type — defaults to salesy. User can request follower-growth for variety.
  • B-roll density — High / Medium (default) / Low / None. See Glossary.
  • Visual direction — model look, environment, lighting, camera style, color palette, aesthetic vibe, production level.

If the user supplies a Researcher creative template, the visual direction is pre-filled from it.

The Manager summarizes the full brief back to the user for confirmation before moving on.

Stage 2 — Script Creation

Agent: Script Writer (Mode 5 for new scripts, Mode 1 for adaptations).

The Manager delegates to the Script Writer with the brief, audience, and any pattern preferences. The Script Writer returns a script. The Manager presents it to the user for approval.

The approved script text and any compliance notes are passed forward to Stage 3.

Stage 3 — Visual Storyboard

Agent: Visual Planner.

The Manager passes the full approved script text plus visual direction, B-roll density, product info, target audience, sales channel, and any creative template.

The Visual Planner returns two artifacts:

  • A user summary (1-2 sentences per scene) — presented to the user for approval.
  • An internal detail version — used by the Manager when delegating to the Image Prompter and Veo Prompter in Stage 4.

User approval here unlocks Stage 4. After this, generation is automatic.

Stage 4 — PatchWork File

Automatic — no per-prompt approval. Once the Video Plan is approved, this stage runs in one pass.

  1. Image prompts — delegated to the Image Prompter. One prompt per shared reference image group (speaking scenes). B-roll gets one prompt per clip.
  2. Speaking video prompts — delegated to the Veo Prompter. Uses the universal talking head template; only the dialogue line changes per scene.
  3. B-roll video prompts — delegated to the Veo Prompter. Simple natural language, no dialogue, ambient audio only. Note: B-roll prompts run manually, not inside the .nbflow. They live in Assets/{workflow}/broll/.
  4. The .nbflow itself — delegated to the PatchWork Importer. Wires up image gens → scene Veo3 nodes → Approval gates.
  5. Save outputs — .nbflow lands in projects/{month}/{brand}/{growth|sales}/{testing|approved}/{workflow}-{version}.nbflow.
  6. Reference material checklist — listed for the user (avatar photo, product photo, etc.).

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

Stage 5 — Generation

Agent: Generation Runner.

Two execution paths:

A. Headless via workflow-runner.js (default). Calls the G-Labs API directly, uploads reference images to R2, exports a completed .nbflow with all generated URLs baked in. Fastest, no browser.

B. Playwright-driven via the PatchWork web UI. Drives https://patchwork-33m.pages.dev/ in a Playwright browser. Useful for visual verification or when the UI flow handles cases the headless runner doesn't.

The runner:

  • Enforces outputCount=4 on every gen node (4 candidates per row).
  • Uploads reference images to R2.
  • Wipes stale cache.
  • Runs all generation nodes with default concurrency 3 (cap 5).
  • Retries failed nodes up to 3 times.
  • Visually reviews every generated image for AI tells (extra fingers, garbled limbs, melted faces).
  • Auto-reruns flagged images via regen-nodes.py, up to 3 attempts per node.
  • Exports a -generated.nbflow to projects/{month}/{brand}/Generations/.

The user opens the exported file in PatchWork to review.

Stage 6 — Delivery & Feedback

The Manager presents the deliverables:

PIPELINE COMPLETE

VIDEO PLAN: [Product] | [N] scenes + [N] B-roll ([density] mode) | [Xs] total
PATCHWORK FILE: [filename].nbflow
GENERATION REPORT:
  Images: [X] gen nodes × 4 candidates each = [Y] stills, [Z] failed
  Videos: [X] gen nodes × 4 candidates each = [Y] clips, [Z] failed
REFERENCE MATERIAL NEEDED:
- Avatar image
- Product image
- [...]
NEXT STEPS:
1. Review and Send video approvals in PatchWork
2. Rerun any failed nodes
3. Export final videos from PatchWork

The feedback loop offers:

  • Script changes → back to Stage 2.
  • Storyboard changes → back to Stage 3, regenerates Stage 4.
  • Specific prompt adjustments → edit individual prompts, regenerate the .nbflow.
  • Rerun failed generations → back to Stage 5.
  • Create a variant (Mode 4 on the script). See Variants & Versioning.
  • Translate (Mode 2 on the script).