How to Make AI Videos That Don't Feel Like AI Slop

Overview

The video explains a “no-AI-slop” workflow for producing AI-generated videos using a local, folder-based pipeline that mimics classic animation studio production:

sound → timing → storyboard → animation → render

A key principle is stage-by-stage checking, so errors don’t compound into finished-video mistakes.

---

Core Technological / Product Concepts

• Folder-based AI automation pipeline

  • Everything needed is contained within one organized folder on the creator’s computer (scripts, markdown files, and stage-specific subfolders).

• AI agent orchestration via files

  • A top-level map file tells the AI what each stage is and where to find everything.

• Stage-by-stage validation (quality control)

  • After every stage, a check happens before moving on, preventing “bad takes” from becoming final outputs.

• Use of external AI tools for specific tasks

  • 11Labs for voice generation (script → spoken audio)
  • Whisper (run on a graphics card) for timed transcription
  • Claude for drawing/scene generation and QC based on generated outputs (e.g., fixing overlaps/cropping)
  • Metricool for scheduling distribution (connection included inside the same folder/workspace)

---

Pipeline (“Five Stages”) Description

Each stage is implemented as its own folder containing:

• What the stage reads
• What it creates
• Who checks it before the next stage runs

1. Voice Stage

• Create director notes via Claude
  • Based on target audience, format, metaphor/style, and the voice to use.
• Generate narration by sending the script to 11Labs.
• Iterative correction
  • If parts sound wrong, re-generate only those pieces.
• Manual QC
  • The creator listens to the full recording before continuing.

2. Words Stage (“Exposure Sheet”)

• Run Whisper to transcribe the narration with exact start/end timestamps.
• Compare transcript vs. the original script and mark differences.
• Manual QC
  • Fix misspelled names based on the transcript.

3. Storyboard Stage (Text-Based Timing)

• Create a plain text storyboard file containing beats tied to specific transcript phrases.
• Run a “cues” command to extract those phrases + timestamps into a table resembling a reel.
• Rationale:
  • Correcting text/beat timing is cheaper than correcting finished animation.

4. Animation Stage

• Claude writes the scene as code
  • Uses reusable “kits” (characters, rooms, UI/screen elements, etc.).
• “Stills” command grabs frames to form a contact sheet (like pencil tests).
• QC via Claude
  • Checks contact sheets for visual issues (overlaps/cutoffs).
• Studio-style QC
  • The creator watches with audio (“sweat box”) and feeds timing notes back into the storyboard.

5. Render Stage

• A command renders frame-by-frame in a browser environment.
• Mixes:
  • voice + music + sound effects
• Produces a smaller phone-optimized copy.
• Adds:
  • descriptions / thumbnails / chapters
• Note:
  • Title/thumbnail selection is kept human, not fully automated.

---

How to Apply This Yourself (Tutorial-Style)

• Don’t start with video—identify any repetitive workflow
  • Examples: newsletters, client reports, onboarding, etc.
• Define stages where each stage hands off to the next.
• Make each stage a folder with a small spec file (reads/creates/checker).
• Add a top-level file that tells an AI how to traverse the stages.
• Emphasize:
  • checking is the most important part
  • Human/AI review between stages prevents cumulative quality failures.

---

Main Speakers / Sources

• Speaker

  • The video creator (author of the workflow)
  • Notes that the narration voice is a cloned voice reading a script stored in the folder.

• AI tools referenced

  • Claude, Whisper, 11Labs

• Other product referenced

  • Metricool