shorts-trends 5.1.0rc2

YouTube Shorts trend analyzer with local dashboard

YouTube Shorts Trend Analyzer

Local Python pipeline that ingests YouTube Shorts data, identifies early viral candidates, scores them with AI-driven content evaluation, and surfaces repeatable viral formats.

Quickstart

python -m venv .venv
.venv\Scripts\activate         # Windows
pip install -r requirements.txt
python main.py --input data/sample_videos.csv

Default scorer is MockScorer (no API key needed) so the pipeline runs out of the box.

AI Features (Claude Code)

The --script, --ideas, and SCORER=claude features authenticate via a signed-in Claude Code CLI session.

Prerequisites

1. Install the Claude Code CLI:

   npm install -g @anthropic-ai/claude-code
   

2. Sign in with your Claude Max plan:

   claude login
   

   Complete the OAuth flow in the browser using the account that holds your Max plan subscription.

3. Verify your sign-in:

   claude /status
   

Running AI features

python main.py --input data/sample_videos.csv --script   # generates 3 production-ready scripts
python main.py --input data/sample_videos.csv --ideas    # generates 2-3 video ideas

All AI usage is billed against your Claude Max plan. No additional API key is required.

Note: API-key fallback (ANTHROPIC_API_KEY-based auth) is NOT supported in this version. Future versions may revisit this if there is demand.

The YOUTUBE_API_KEY environment variable is only needed for --query and --channel (live YouTube fetching), not for AI features.

Async usage (FastAPI)

All three generators expose native async methods (ascore, agenerate) usable directly from async hosts without RuntimeError:

from fastapi import FastAPI
from ai_analysis.claude_scorer import ClaudeScorer
from output.ideas import IdeaGenerator

app = FastAPI()
scorer = ClaudeScorer()
idea_gen = IdeaGenerator()

@app.post("/analyze")
async def analyze(videos: list[dict]):
    # Convert request payload into Video objects (omitted)
    scored = await scorer.ascore(video_list)
    ideas = await idea_gen.agenerate(
        archetypes=detected_archetypes,
        top_videos=scored,
    )
    return {"scores": scored, "ideas": ideas}

The sync entry methods (score_batch, generate) are preserved as thin wrappers and remain safe to call from any host — including async ones — via internal sync-from-async dispatch.

Pipeline

CSV input
   │
   ▼
data_collection      → list[Video]
   │
   ▼
feature_extraction   → metrics: views/hour, age_hours
   │
   ▼
ai_analysis          → hook_strength, curiosity_gap, viral_potential (0-10 each)
   │
   ▼
scoring              → composite 0-100
   │
   ▼
output               → ranked table + pattern report

Modules

Module	Purpose
config/	Weights, thresholds, model settings
data_collection/	CSV loader (v1), YouTube API client (v2)
feature_extraction/	Numeric metrics (views/hour, age)
ai_analysis/	Pluggable LLM scorers (Mock / Claude)
scoring/	Combine metrics + AI scores into composite
output/	Rank, print, extract patterns

Production Pipeline (v5.0)

The v5.0 production pipeline turns a generated script into an upload-ready YouTube Shorts bundle:

shorts-trends produce --from-script <script.json> [--variant N] [--length 15s|30s|60s] [--out <dir>]

The bundle directory <out>/<run_id>/ contains:

• <run_id>.mp4 — final 1080×1920 stitched short
• metadata.json — title (≤95 chars), description (first line contains #Shorts), tags (within YouTube's 500-char budget), category_id: 22, privacy_status: "private", thumbnail_kind: "auto"
• <run_id>_thumbnail.png — auto-extracted midpoint frame
• credits.json — Pexels attribution list
• <run_id>.srt — sidecar captions (YouTube CC parity)
• upload_instructions.txt — step-by-step manual upload guide

Manual upload boundary

v5.0 produces an upload-ready bundle. Manual upload via studio.youtube.com is the supported path. Programmatic upload (--upload flag, OAuth flow, credentials cache) is on the v5.1 roadmap (UPLOAD-V51-01).

Following each shorts-trends produce run, open upload_instructions.txt in the bundle directory for the exact YouTube Studio steps. The metadata.json fields are designed to be copy-pasted directly into the YouTube Studio upload form.

Roadmap

• v1 (this scaffold): CSV → mock scoring → ranked output
• v2: Plug in Claude (claude-agent-sdk) — auth delegates to a signed-in Claude Code CLI session.
• v3: YouTube Data API v3 collector replacing CSV
• v4: Scheduling (cron / Windows Task Scheduler) + persistence (SQLite)