Empirical Workflow State Progression

PROD: PASS AUD: PENDING
1. ONE QUESTION
2. Notebook Research (29 sources)
3. Video Generation R0
4. AI Media QA
REVISE (5 Defects)
5. Evidence QA
CORRECTIONS
6. AI Autonomous Repair Prompt
7. Video Generation R1
8. Production Gate
PASS (通过)
9. YouTube Published Artifacts
10. Audience Gate
PENDING (观察中)
Current Empirical Verdict

Automation candidate validated once; repeatability pending.

REPEATABILITY: UNVERIFIED AUTOMATABLE: NOT YET
Real Workflow Test #Q03 Observed: 2026-09-15
Production Gate: PASS Audience Gate: PENDING

Can AI Turn One Prompt Into a Publishable YouTube Video?

Q03 · Real Workflow Test — Generation → QA → Repair → Publish

Primary Engine
Official Cloud Pipeline
Human Video Editing
0 min
Zero Manual Timeline Editing
Production Gate
PASS (通过)
Audio sync, pacing & encoding
Audience Gate
PENDING (Pending)
Live YouTube signals tracked

1. The Question & Human Participation Boundary

Can a normal creator give AI one topic, avoid manual research and video editing, and still receive something worth publishing on YouTube?

Minimize human production work while preserving enough factual reliability, narrative quality, and visual coherence to cross a real publication gate.

Permitted Human Role (Observed)

  • · Define the topic
  • · Provide workflow prompts
  • · Initiate Notebook generation
  • · Review the final candidate
  • · Make the final publish/reject decision
  • · Upload the accepted outputs

Strictly Excluded Human Actions

  • · Manually research the topic
  • · Write the final script
  • · Edit the video timeline
  • · Replace B-roll manually
  • · Rewrite factual claims directly inside the video
  • · Manually repair audio or animation

Structured Human Effort Ledger

Human Research Minutes 0 min
Human Script Editing Minutes 0 min
Human Timeline Editing Minutes 0 min
Human Prompting Occurred (orchestration)
Final Human Decision Required (Human Gate)
This distinction matters. The task is testing AI production, not whether a human editor can rescue AI output.

2. Generation R0 Autopsy — 5 Material Defects

STATE: FAILED_QA

Gemini Notebook produced a full long-form explainer and vertical Short without manual video editing, but failed internal publication review due to 5 material defects.

1. Incorrect Production Provenance

The video claimed generation by Remotion and ElevenLabs. In physical reality, it was generated primarily through Gemini Notebook.

2. False Precision

Contained simulated values such as "90% AI / 10% Human", estimated review minutes, and hypothetical labor costs presented as empirical observations.

3. YouTube Policy Oversimplification

Implied that YouTube penalizes AI-generated video simply because AI is involved, rather than focusing on repetitive, template-driven, or low-value spam content.

4. Synthetic-Media Disclosure Overgeneralized

Implied that automated uploads must universally set status.containsSyntheticMedia = true, conflating realistic altered reality with stylized animations.

5. Visual Repetition

Over-relied on uniform white grid backgrounds, hand-drawn icons, and infographic cards throughout the long-form runtime.

3. Published Artifacts — YouTube Verification

GATE: PRODUCTION PASS

Empirical Boundary Note: The public links below represent the final published candidate artifacts (16:9 Long-form Video and 9:16 Vertical Short) generated through the autonomous pipeline, both having passed the Production Gate. The R0 to R1 failure and repair progression is documented via the structured QA audit matrix and regeneration ledger below.

16:9 Long-form Video Production Gate: PASS

Can AI Make a Publishable YouTube Video From One Prompt?

Full-length candidate video generated through NotebookLM research and structured prompt orchestration. Passed audio sync, narrative arc, and broadcast encoding gates.

9:16 Vertical Short Production Gate: PASS

How AI Repairs Its Own Video Mistakes

Vertical short-form candidate distilled from the core research brief. Passed rapid hook, vertical framing, and pacing gates.

4. Independent AI QA Architecture & 3-Column Audit Matrix

Instead of repairing the video manually, the experiment introduced an independent decomposed AI review layer. Attempting to make one Gemini instance simultaneously audit facts, check media, score pacing, and verify platform compliance crashed with timeouts. Automation requires cognitive role decomposition.

Decomposed QA Pipeline: Video → Media QA → Claim Inventory; Claims + Research → Evidence QA → Safe Claim Ledger; Media QA + Evidence QA → Repair Instructions

Retained Facts (KEEP)

  • AI can automate substantial portions of research, scripting, narration and assembly.
  • Fully autonomous production still creates quality-control problems.
  • Modular pipelines offer more control but introduce engineering overhead.
  • Human repair should be counted as part of the real production cost.

Refined Framing (QUALIFY)

  • YouTube monetization policy claims.
  • Synthetic-media disclosure requirements.
  • Tool-specific performance and duration limits.
  • Cost comparisons between all-in-one and modular systems.

Hallucinations Removed (DELETE)

  • Unsupported percentages and simulated stats.
  • Hypothetical labor numbers presented as observations.
  • Universal claims about AI demonetization.
  • Claims that this video was generated by infrastructure that had not actually been used.

5. Generation R1 — 8 Autonomous Improvements

STATE: PASSED_QA

QA findings were synthesized into a structured repair prompt without manual script rewriting or timeline editing. Gemini Notebook regenerated both videos, resolving all 5 primary defects.

1 Corrected production provenance to honest toolchain disclosure
2 Removed false-precision metrics and simulated labor costs
3 Accurately distinguished YouTube anti-spam policies from AI creation
4 Clarified exact statutory boundaries of synthetic media disclosure
5 Avoided claims of universal optimality
6 Incorporated the actual experiment failure and repair into the narrative
7 Significantly diversified visual pacing and infographic layout
8 Strictly reserved judgment pending post-release audience verification

6. Multi-Gate Outcome & Effective Video Cost Model

Production Gate

PASS

Video encoding, audio sync, narrative arc, and broadcast standards were achieved without manual script or timeline editing.

Audience Gate

PENDING

Published artifacts are live on YouTube. Objective retention, CTR, average view duration, and subscriber signals are currently being monitored.

9bests Effective Video Cost Model

Framework, not measured total
Effective Video Cost = AI Cost + Human Repair Cost + Failed Generation Cost

Framework, not measured total. Token and API billing were not isolated under measured sub-accounting, so no dollar amounts are claimed.

Epistemic Defense: What Was NOT Proven

  • Did NOT prove Gemini Notebook / NotebookLM is the best AI video system.
  • Did NOT prove this workflow is universally optimal across genres.
  • Did NOT prove AI-generated videos will achieve strong viewer retention.
  • Did NOT prove fully autonomous channels are commercially viable.
  • Did NOT prove eliminating human review improves video quality.

7. Complete Task State Matrix

Dimension / Stage Empirical Status
Topic supplied by human Yes
Manual research No (0 min)
Manual script writing No (0 min)
Manual timeline editing No (0 min)
AI Deep Research Yes (29 sources)
AI long-form generation Yes
AI Short generation Yes
Independent AI Media QA Yes
Independent Evidence QA Yes
First generation (R0) REVISE (5 QA defects)
AI regeneration (R1) Yes (8 improvements)
Production Gate PASS
Audience Gate PENDING
Repeatability verified UNVERIFIED
Fully automatable pipeline NOT YET
Retest required Yes

8. Retest Plan & Verification Roadmap

REPEATABILITY: UNVERIFIED AUTOMATABLE: NOT YET STATUS: PLANNED

Verification Roadmap Questions

  • ? Does the workflow cross the publication gate again on a completely different topic?
  • ? How many automated repair rounds are required?
  • ? Does Human Repair remain near zero minutes?
  • ? Does the visual format become repetitive across videos?
  • ? Does the real audience watch beyond the 30-second retention cliff?
  • ? Does effective cost improve with repeated workflow usage?
Tested Engine Profile
NotebookLM (Gemini)

Google's AI research assistant that synthesizes information from your documents into insights and audio summaries

Catalog Pricing: Free
Inspect Tool Details & Pricing