The Retention Script Engine, Hooks, and Word Budgets

In the world of YouTube algorithms, retention is the only metric that cannot be faked.

A video with an exceptional thumbnail but poor retention will experience initial clicks, suffer immediate audience drop-off, and be discarded by the recommendation engine within 48 hours.

Conversely, a video that maintains a 50%+ Average Percentage Viewed (APV) on a 15-minute runtime triggers aggressive, continuous Browse distribution for months or years.

01. The 0–3 Second Hook Formula

The first 3 seconds of your video decide whether a viewer watches until the end or clicks back to their feed.

Three Banned Hook Openers:

• The channel logo bumper with generic EDM music.
• “Hey guys, welcome back to the channel, make sure to like and subscribe.”
• “In today’s video, we are going to be talking about the fascinating history of…”

The High-Tension Hook Blueprint: Open in media res—in the middle of the most intense, unresolved paradox of your story:

“In November of 1968, a Soviet nuclear submarine vanished 16,000 feet beneath the Pacific. What the US Navy discovered on the seabed remained classified for thirty years… because the crew wasn’t dead when the wreckage was found.”

This introduces an urgent psychological question that can only be answered by staying until the climax.

Tool Walkthrough: Browse our curated Hook Library. It contains hundreds of platform-tested curiosity openers categorized for long-form documentaries, YouTube Shorts, and viral essays.

02. The Two-Phase Script Pipeline

Never attempt to generate a 2,000-word script in a single AI prompt. Single-shot prompts lose structural pacing, repeat metaphors, and drift into shallow generalities.

Instead, top creators use a Two-Phase Hierarchical Engine:

1. Phase 1: Structural Beat Outline (Claude Haiku): A fast, structured LLM call maps the video into rigid chapter roles: Hook, Setup, Escalation, Turn, Climax, and CTA. Each beat is assigned an exact narrative role and word budget.
2. Phase 2: Parallel Beat Expansion (Claude Sonnet 3.7): Each beat is expanded independently with strict duration constraints. By caching system prompts and reference context via Anthropic’s ephemeral prompt caching (cache_control: ephemeral), multi-beat script generation and inline revisions cost 90% less in input tokens.

Tool Walkthrough: Before generating audio, verify your word count with our Script to Time Calculator. Natural documentary narration runs at exactly 140 to 145 spoken words per minute. If you want a crisp 12-minute video, your script must measure between 1,680 and 1,740 words.

03. Grounding Packs to Kill Hallucinations

Viewers immediately recognize fabricated historical dates, false statistics, or artificial dialogue. Negative comments pointing out factual errors destroy audience trust.

Before expanding beats, build a Grounding Reference Pack:

• Scrape 2 to 3 verified research articles or historical archives into plain text.
• Extract raw transcripts from authoritative documentary coverage.
• Inject this text as a cached reference block so the LLM is strictly constrained to verifiable historical facts.

Tool Walkthrough: Don’t build story structures from scratch. Choose an battle-tested architecture from our Template Vault (Documentary, 10 Secrets Listicle, Timeline Investigation, or Contrarian Explainer) and launch it directly into Faceless Studio with pre-configured beats.