Automating YouTube Shorts Without Coding: A 9-Step Pipeline Built with Claude Code and Gemini

AI creator @Che0dan has published a comprehensive 9-step roadmap demonstrating how non-programmers can automate the entire YouTube Shorts production workflow using modern AI coding agents such as Claude Code and OpenAI Codex.

Image source: @Che0dan via X

Rather than relying on vague, one-off prompts, the guide presents an end-to-end software architecture built through "vibe coding." By separating computationally heavy video processing on the local machine from high-leverage cloud AI models, creators can build a sustainable pipeline that covers everything from raw footage segmentation and fact verification to multi-variant scriptwriting, voiceover synthesis, and an in-browser human review dashboard.

Core Architecture: The 4-Phase Loop and AI Tool Stack

The entire automation system follows a structured four-stage data cycle:

1. Read: Analyze the source video, split it into distinct visual scenes, and record timecodes and scene features.
2. Write: Extract verifiable facts from the video description or source notes, then generate titles, subtitles, and three distinct script variations (Type A, B, and C).
3. Make: Synthesize narration using TTS, then combine video clips, subtitles, and audio via local ffmpeg commands into three vertical 1080x1920 Shorts.
4. Verify: Display the rendered outputs in a local FastAPI web dashboard for human review and final approval.

The technical stack incorporates the following tools and models:

• Development & Vibe Coding: Claude 5.5 Opus, Claude 5.5 Sonnet, OpenAI Codex
• Information Extraction & Voice Synthesis: Google Gemini 3.8 Flash, Gemini 3.8 Flash TTS (local alternatives: Supertonic-3, Qwen3-TTS)
• Auxiliary Verification: Clef-omni on Cloudflare Workers AI (a Jev-compatible evaluation model at $0.15 per million input tokens)
• Local Media & App Framework: Python, PySceneDetect, ffmpeg, FastAPI

The 9-Step Implementation Workflow (Steps 0 to 8)

The step-by-step pipeline outlined by @Che0dan moves methodically from foundational scaffolding to automated safeguards.

Step 0: Project Specification & Git Setup

Before writing code, instruct the AI coding agent to generate a comprehensive PLAN.md outlining the architecture, file layout, and target criteria. To prevent accidental leaks of sensitive keys and large media assets, configure .gitignore to exclude .env, .venv, input/, output/, data/, models/, and backup/ before making the initial Git commit.

Step 1: Baseline Pipeline Verification

Establish a minimal working baseline using Python and system ffmpeg. Test that the first 30 seconds of a raw video (input/source.mp4) correctly crops and renders into a 1080x1920, 30fps, H.264 vertical video (output/step1.mp4).

Step 2: Scene Segmentation with PySceneDetect

Integrate PySceneDetect to automatically detect visual transitions and divide the footage into scenes. Extract two representative keyframe stills per scene and log the precise start/end timecodes and metadata into scenes.json.

Step 3: Fact Extraction & Grounded Verification with Gemini

Call the Gemini API to parse the source text or video description for key thematic metadata. To eliminate AI hallucination, enforce strict code-level verification so that only facts with verifiable evidence in the source text are saved into data/facts.json. Unsupported claims are blocked from generation and flagged for manual review.

Step 4: Multi-Style Script Generation

Using the verified facts in data/facts.json, prompt the model to produce three distinct script formats simultaneously:

• Type A (Subtitle-Driven): Fast-paced pacing focused on punchy, text-heavy captions.
• Type B (Hook & Twist): High-retention opening hooks with dynamic structural shifts.
• Type C (Narrative Explainer): Calm, informative storytelling emphasizing clear context.

Step 5: Voiceover Synthesis & STT Reverse Check

Convert the generated scripts into spoken audio using Gemini 3.8 Flash TTS (or local models like Supertonic-3 and Qwen3-TTS) and store them in audio/. To catch mispronunciations, dropped words, or audio artifacts, run the audio through an STT (Speech-to-Text) pass and compare the transcription back against the original script text.

Step 6: Local ffmpeg Composition & Quality Check

Assemble the segmented video clips, subtitles, and synthesized audio tracks using local ffmpeg commands. Render the three final 1080x1920 vertical video files and inspect the rendered frames to verify visual output.

Step 7: FastAPI In-Browser Review Dashboard

Build a clean browser interface using FastAPI so non-technical users can operate the tool effortlessly. The dashboard enables local file uploading, displays live pipeline progress, provides side-by-side video playback of all three rendered variations, and includes an explicit approval action.

Step 8: Error Handling & Operational Safeguards

Harden the pipeline for production reliability. Implement execution locks to prevent overlapping duplicate runs, resume logic to restart from the last successful stage after a failure, API rate-limit handlers, automatic project backup and restore capabilities, and automated smoke test suites.

Production Caveats and Token Budget Management

@Che0dan highlights several critical engineering and operational principles when running this automated pipeline:

• Strict API Key Hygiene: Never paste paid API tokens directly into chat prompts. Always store credentials in a local .env file and ensure it is shielded by .gitignore.
• Decoupling Local Compute from Cloud AI: Heavy media manipulation—such as video re-encoding, cropping, and subtitle burning—must run on the local host via ffmpeg. Relying on cloud AI agents to process raw video frames burns token limits within minutes. Restricting cloud models to text analysis, fact extraction, and script generation avoids exhausting AI usage limits while processing video locally.
• Source Footage Rights: Avoid scraping unauthorized third-party Shorts from social platforms. Use licensed stock, permission-granted assets, or self-recorded footage as the primary video source.
• Human-in-the-Loop Verification: Automated generation should never bypass human editorial judgment. A human operator must always review and listen to the final rendered files before publishing to verify factual accuracy and visual appeal.

Original source

• Original X Thread: @Che0dan on X
• Detailed Article: Original X Article