YouTube Growth Guides

YouTube audience retention is the share of a video that the average viewer watches, and the Studio graph shows the moment each group left. This guide explains how to read that graph, then walks through the script fixes that improve audience retention on a video you have not recorded yet.

36 copy-paste YouTube hooks across 9 patterns: curiosity gap, contrarian, proof, story, direct value, question, preview, demonstration, and stakes. Each comes with the psychology behind why it works and three variations to adapt. The opening seconds are where a viewer decides, which is why the hook is the first thing we script.

The best YouTube script structure is Hook, Build, Payoff: a 0-30 second hook, a body that changes something every 30-60 seconds to reset attention, and a payoff that lands the aha moment. It is the framework high-retention creators reuse on every video, and it works for any content type.

The first 30 seconds are where a YouTube opening does its work: the phases below are Prepublish's guidance for that window, with hook patterns by niche and four before/after rewrites. We do not cite a YouTube rule for them, and we do not publish a retention benchmark for your own videos.

YouTube script analysis is the practice of evaluating a video script before recording, to map where the draft carries attention risk while it is still text. This page defines the practice: what it measures, what a tool can and cannot see from a script, and how it differs from a grammar checker or an SEO suite.

A 27-step checklist covering script validation through upload optimization, organized by production phase with a pass condition for every step. The step most creators skip: checking the script for structural problems before recording.

The four mechanisms this guide is built on: the Hook Trident, the SPEC outline method, the 300-word section rule, and the adversative transition technique. The generic write-a-script walkthrough is compressed into one summary at the end.

6 specific narrative structures used by high-retention YouTube videos, broken down beat by beat with timing guidelines and retention data. Includes the Curiosity Loop, the Transformation Arc, the Problem Stack, the Expert Contrast, the Ticking Clock, and the Reveal Ladder.

Pattern interrupts are deliberate changes in pace, tone, format, or visual presentation that reset viewer attention. This playbook covers 14 specific techniques with exact timing formulas, category-specific strategies, and 4 popular interrupts that actually hurt retention.

YouTube retention drops follow 8 identifiable patterns, each with a specific cause and fix. Learn to read your retention graph like a diagnostic tool and trace each drop to the exact script-level problem that caused it.

Viewer retention is driven by specific cognitive mechanisms. Understanding 10 psychological principles gives you a framework for writing scripts that work with the brain's attention systems instead of against them.

Tutorials are YouTube's largest content category by volume, with a unique retention challenge: viewers come with a specific goal and leave the moment they believe they've achieved it or concluded they won't. This guide covers the 6 structural mistakes that cause tutorial viewers to leave.

The best YouTube script tool depends on the job you need done. Use 1of10 for topic research, Subscribr for AI script drafting, and Prepublish for retention auditing before you film. ChatGPT works for flexible drafting, Spotter Studio handles packaging, and Google Docs remains the standard for collaborative writing. This guide breaks down seven tools by their actual function, not a ranked hype list.

On a faceless channel, the script is the product. You have no personality on camera to carry weak writing. This guide covers how to structure faceless scripts, pick the right narration pace, and edit AI drafts so they sound human. Learn the specific retention traps that kill faceless videos and how to fix them before you pay for a voiceover.

There is no official retention benchmark for YouTube Shorts, so the first seconds are best judged against your own recent Shorts. Shorts percentage viewed can even exceed 100 percent because loops and replays count toward watch time in YouTube Studio. This page is Prepublish's guidance: your first frame is the hook, you need a beat every few seconds, and your ending should feed back into the beginning.

Five manual tests to run before you record, covering the hook, length, payoffs, title match, and cuts, followed by an automated check for what self-review keeps missing.

An AI-written script can read smoothly and still fail on camera. This guide covers the six checks to run before you record or voice it, what YouTube's inauthentic content policy says about AI-assisted scripts, a manual checklist, and where a script checker fits.

Feedback on a published video is an autopsy. Feedback on a script is a rescue. This guide maps where creators actually get script feedback, how to ask so strangers answer, and how to decide which notes deserve a rewrite.

A script analyzer reads your script before you record and flags the passages most likely to lose attention. Only a handful of tools genuinely do this. This comparison covers all of them, including ours, and is honest about which one fits which creator, starting with a free check that needs no account.

Watch time is views multiplied by average view duration, and almost all durable growth comes from the duration side. This guide covers the seven script-level fixes that raise it, the video-length trap, and the realistic math behind 4,000 watch hours for monetization.

On channels between 1,000 and 20,000 subscribers, hits and flops open nearly identically. A preregistered 941-video study found none of the three hook patterns that separate hits from flops on larger channels. Here is what that means for where your effort should go.

The median flop does about a third of its channel's normal views, and on small channels the opening is rarely the reason. Two studies of 1,290 videos suggest a checking order: topic first, packaging second, script third.

"Talk faster, cut every pause" is default YouTube advice. Across 1,290 videos in two studies, hits and flops speak at nearly identical pace: 181 vs 180 words per minute on larger channels, 169 vs 170 on small ones.

By word 15, both hits and flops have made a point. The difference is what the point is made of. Data from two studies covering 1,290 videos.

"Hey guys, welcome back to the channel" is the most-mocked opening on YouTube. The data mostly agrees, with more nuance than the mockery: greetings leaned toward flops in both studies, and it is the only opening signal that repeated.

Audit a YouTube script in Claude by running four checks in order: score the opening, audit the full draft, run the policy and reused-content checks, then fix the top passage and re-score. Claude calls published checks and reports their named fields, and none of them predicts performance.

YouTube auto captions roll, so each cue repeats the lines already on screen and appends the one being spoken, and a naive extraction reads every sentence two or three times. The fix is one rule: drop any line identical to the line emitted immediately before it, after stripping inline tags and collapsing whitespace.

Prepublish serves a YouTube script audit over MCP at https://mcp.prepublish.ai, and Claude Code connects to it with one command. The server needs no account or API key, exposes six tools, and answers each call with scores, flagged passages and rewrites for a script you have not recorded yet.

The Prepublish channel skill generator reads the 12 most recent long-form uploads of a public channel, pulls the original-language captions from at least four of them, and writes a four-file Agent Skill bundle in about six minutes. The bundle is one SKILL.md brief plus three reference files of verbatim evidence: hooks, titles, and transcript excerpts.

An agent skill is a folder whose name matches the name field in its SKILL.md, and each client reads it from a directory that client owns. Claude Code reads project skills from .claude/skills/ and personal skills from ~/.claude/skills/, while Cursor reads .agents/skills/ or .cursor/skills/ in a project and the same two names under your home directory.

An auto-translated YouTube caption is not a caption track. The player builds one by adding tlang to a real track's URL, so the request keeps lang=en and returns Turkish. The spoken language lives in audioTracks[defaultAudioTrackIndex], and that field, not the viewer's selection, is what a caller should select on.

Prepublish serves a remote MCP server at https://prepublish.ai/mcp from one Next.js App Router route handler and no server-side SDK. A GET to that URL returned a 1,749-byte JSON server card on 19 September 2026, and the handler, the caps, the bridge and the conformance suite are quoted here from the files that run it.

SKILL.md holds instructions an agent loads when a task matches its description, CLAUDE.md and AGENTS.md hold project context loaded at session start, a system prompt travels with every request, and an MCP server exposes tools the model can call. Decide by how often the context must be present and whether the agent has to act on something outside the conversation.

Adjectives do not transfer to a writing assistant, and quoted examples do. The rules worth handing over are already in your transcripts: how an opening starts, how long the sentences run, which words the channel avoids, and when the point arrives.

YouTube's timedtext endpoint answers HTTP 200 with a zero-byte body when the request does not carry a proof-of-origin token minted inside the player. The token is absent from the page HTML and from ytInitialPlayerResponse, and every route that works either replays the player's own request or presents a browser TLS fingerprint.

The yt-dlp bot check on a server comes from the egress IP, while empty caption bodies come from a missing browser TLS fingerprint. Those are two faults with opposite fixes, and a canary that reports them apart is what keeps a deployment honest.

You can turn a YouTube video into a Markdown file three ways, and they fail differently. The transcript panel gives you timestamped lines and no metadata, yt-dlp gives you the caption track you ask for, and a browser extension writes the metadata header the other two leave out.

Claude Desktop reaches the Prepublish YouTube script audit two ways: a custom connector under Customize then Connectors, or npx -y prepublish-mcp in claude_desktop_config.json. The connector needs no file and no key, and the file route is the one that runs the server on your own machine.

Prepublish is in ChatGPT's plugin directory, so the YouTube script audit installs in three steps: open Plugins in the sidebar, search Prepublish, select Install plugin. No developer mode, no key, and it works across plans on web, desktop and mobile. Developer mode stays the route for an unlisted server or a workspace that blocks public plugins.

Cursor adds the Prepublish YouTube script audit with one entry in mcp.json, either the project file or the global one, or in one click from Customize. A stale entry in the project file overrides the global one silently, which is the failure to check first when the server looks wrong.

Cline adds the Prepublish YouTube script audit from its Remote Servers tab or from the MCP settings JSON, and one field decides whether it connects at all: the transport type has to say streamableHttp, because leaving it out defaults to the legacy SSE transport that this server does not serve.