---
name: youtube-playbook
description: Make YouTube videos that get clicked and watched to the end. Use when the user wants to research video ideas, analyze competitors or a niche, plan, script, title, or thumbnail a YouTube video, improve retention or pacing, diagnose an underperforming video, or build a channel strategy. Covers niche research with scrapers and tools, idea validation, the title/thumbnail-first workflow, story-arc scripting, pacing and pattern interrupts, thumbnail composition and color, and reading retention data. Built from practices tested across channels with over a billion combined views.
---

# The YouTube playbook

This skill runs the workflow behind channels that grew to over a billion combined views.

The order matters more than anything in it: **title and thumbnail come before the video exists.** Most creators shoot first and ask "what's the thumbnail?" at upload time. It works the other way: if there's no clickable title and thumbnail, the video doesn't get made, no matter how good the idea sounds. Everything starts with the click, because a video nobody clicks is a video nobody watches. And before the title comes the research, because the niche's data already knows what gets clicked.

## Phase 0: research the niche with data, not vibes

Before ideas, build a picture of what actually works in the niche. This is agent work: scraping, computing, and pattern-finding that used to take weeks of manual browsing.

**Build the competitor set.** List 10-20 channels serving the same audience: the big ones, the fast-growing small ones, and adjacent niches that share the viewer. Fast-growing small channels matter most, because whatever they're doing is working without the advantage of an existing subscriber base.

**Scrape the niche.** Apify's YouTube scrapers (or the free YouTube Data API if the user prefers to set up a key) can pull, per channel: every video's title, view count, publish date, duration, and thumbnail URL. Pull the last 50-100 videos per competitor channel into one dataset. This is the raw material for everything below.

**Find the outliers.** For each channel, compute its median views over the period, then flag every video that did 3x or more its channel's median. Outliers are the signal: a video that beat its own channel's baseline proves the topic and packaging pulled new interest, independent of channel size. A 50k-subscriber channel with a 2M-view video is telling you more than a 10M-subscriber channel with a 2M-view video. Collect the outliers across all competitors into an outlier list; this list is the niche's proven demand, and video ideas come from it, not from brainstorming in a vacuum.

**Extract the patterns.** Across the outlier list, analyze: title structures (question vs number vs challenge vs versus), title length, which emotional angle (curiosity, fear, aspiration, controversy), thumbnail composition (face vs object vs text, color schemes), and video length. Also pull each outlier's thumbnail URL into a swipe file for reference during thumbnail design; study the patterns, never copy an image.

**Mine the comments.** Scrape the comments of the top outliers (Apify has comment scrapers for this). Comments are free audience research: recurring questions become future video ideas, complaints reveal what's missing in the niche, the exact vocabulary viewers use becomes title language, and highly-upvoted comments show what moments resonated. A comment like "part 2 on X please" appearing across multiple channels is a validated video idea nobody's made yet.

**Check search demand.** vidIQ and TubeBuddy show search volume, competition scores, and related terms for keywords in the niche, and YouTube's own search autocomplete (type the seed term, harvest the suggestions, then add each letter of the alphabet after it) maps what viewers actively look for. Use these for ideas and validation. Don't disappear down the SEO rabbit hole: search is one traffic source, and browse/suggested traffic driven by CTR and retention is usually the bigger one. A keyword score never overrides the question "would my specific viewer click this."

**Maintain the dataset.** Re-scrape monthly. New outliers show where the niche's attention is moving, and a topic whose outliers cluster in the last 60 days is a trend worth riding now.

## Phase 1: validate the idea as a click, not a video

Every idea, whether it came from the outlier list or the user's head, gets pressure-tested as packaging before anything is scripted:

1. **Brainstorm titles in volume.** Ten or more per idea, no filtering during the brainstorm. Big campaigns have tested over a hundred title variations before committing, and a well-chosen title/thumbnail combination can multiply CTR several times over. The winner gets picked by response, not by which one the creator likes.
2. **Run the conversation test.** Imagine how a viewer would describe the video to a friend afterward: "did you see the one where...". If the title can't be remembered, explained, and shared in one casual sentence, keep working. A video that can't be talked about can't spread.
3. **Check the Why and the What.** People come to YouTube for four reasons: to be entertained, to learn, to be inspired, or to relax. The title signals which of those the video satisfies (the Why) through a concrete topic hook (the What). If it doesn't clearly answer "what am I getting and why do I care" in a two-second read, it fails at the moment of decision.
4. **Cross-check against the outlier patterns from Phase 0.** An idea whose shape already appears in the outlier list is validated demand; an idea with no cousin anywhere in the niche's outliers needs a stronger reason to exist.
5. **If no title/thumbnail combination survives, don't make the video.** Move to the next idea. This single rule saves more production time than everything else combined.

## Phase 2: the thumbnail

The image does most of the work. Viewing research on browsing behavior found the artwork takes the large majority of a viewer's attention, with the title getting under two seconds. On modern YouTube it's even tighter: the thumbnail holds for roughly two seconds before autoplay kicks in, so the image has two seconds to earn the click and then the video's first five seconds have to earn it again.

Principles that survive testing:

- **Curiosity earns the click.** The image should make the viewer ask a question only the video answers. The strongest test of a concept: does it open a "how is that possible / what happens next" gap? An image that merely illustrates the topic loses to an image that opens a loop. The reference standard: a character standing somewhere they shouldn't be able to reach, mid-action; the viewer has to know how.
- **Faces work because brains are wired for them.** A close-up face with a genuine, readable emotion connects, and the eyes matter most: bright, visible, aimed at the camera or at the object of interest. Exaggerate the emotion slightly beyond natural; at feed size, subtle reads as blank.
- **Face plus object** when there's a thing to show: the emotional reaction next to the interesting object is the most reliable general-purpose format. **Object-only** when the thing itself is the star: huge, centered, bold, zero ambiguity about where to look.
- **Tell the story without words.** If the image can't communicate wordlessly, add minimal text only to clarify or amplify curiosity, three or four words maximum, never a caption repeating the title. Text competes with the title; the image shouldn't need it.
- **Composition: rule of thirds.** Place the subject on the intersection points of a 3x3 grid; eyes go to those points naturally and balanced images process faster. Break the rule when a shot genuinely works better without it, and fix framing in the crop when needed.
- **Color decides more than people think.** Color carries a huge share of a snap decision. Use complementary colors (opposites on the color wheel) so subject and background separate hard; check the thumbnail against YouTube's own UI (white in light mode, near-black in dark mode) so it doesn't dissolve into either. Each color carries an emotional tone, so pick the palette for the feeling first, visibility second, and keep a consistent accent color across the channel so returning viewers recognize videos at a glance.
- **The format library:** face-first with strong emotion; face plus object; object-only; two faces flanking an object (faces left and right, object center); before/after two-panel; three-panel progression reading left to right; scale and perspective extremes (very big or very small next to a person); action frozen mid-moment ("what happens next"); organized clutter (a patterned mass of one thing with the subject popping out of it); arrows, circles, and pointing to direct the eye, used sparingly. Combine two or three tactics; pick by what the story needs, and check the Phase 0 swipe file for what's overused in the niche, because looking like everyone else's thumbnail is its own failure.
- **Check it at mobile size, in context.** Most viewing is mobile. Screenshot real search results or a real homepage and mock the thumbnail into the lineup to see it as a viewer will: small, surrounded by competitors. Squint-test it. If anything is too small to read or distinguish, redo it. Also run the blink test: glance at it for a fraction of a second, eyes closed, and note what registered first; that's all a scrolling viewer gets.
- **Always produce variants.** Two or three distinct concepts per video, shot in a dedicated photo session planned before filming, multiple angles and expressions per concept, never a screengrab hunted at upload time. Audit each candidate: which is more clickable, does it honestly portray the content, does it create excitement or intrigue, would you click it.
- **Test, don't debate.** YouTube's built-in Test & Compare runs up to three thumbnails against each other on real traffic; TubeBuddy offers A/B tooling as well. The standing rule: two thumbnails tested per video, winner stays. If early CTR is weak, swap the thumbnail before giving up on the video; a thumbnail change can revive a dead video days later.
- **Clickbait calibration:** a dramatic thumbnail is fine if the video delivers exactly what it promises. The line isn't drama, it's deception. Broken promises destroy retention, and retention is what the algorithm actually rewards.

## Phase 3: the title

The title's job is to reinforce the thumbnail and give the click its final push, in under two seconds of reading. Thumbnail and title should never say the same thing; the pair works when each adds context the other lacks.

- Active voice, clear and direct, with natural urgency.
- Speak the audience's exact vocabulary, harvested from the Phase 0 comment mining. A word that means nothing to the specific viewer kills the click, and a word from inside their world signals "this is for you."
- Formats that reliably work: questions, open or closed, amplified with a surprising fact, person, or place; numbered listicles (clear expectation, clear time commitment, memorable); problem plus promised solution; urgency and FOMO framing; a trusted authority name for validation; addressing the viewer directly ("you") so the video feels made for them; honest controversy, drama, or opposition where it exists.
- Capitalize Each Word, and put ALL CAPS on one or two keywords for emphasis; all-caps everywhere reads as yelling.
- Amplifier words (best, worst, insane, ultimate, easy, step by step, in 24 hours) push the keyword over the edge; place them on the keyword, don't scatter them.
- Trending and topical framing when the niche has a live moment: riding a current event or a niche controversy multiplies reach while it's hot, and the monthly re-scrape shows what's hot.
- Length: front-load the payload. Titles truncate in some surfaces around 50-60 characters, so the words that sell the click go first.
- Generate 10+, shortlist 3 by the conversation test, then let CTR data pick the final winner, and be ready to retitle a live video whose impressions are high but CTR is low.

## Phase 4: the script

Every great video follows the same story arc, whether it's entertainment, education, or a product video. Script to this shape:

1. **The hook.** The first five seconds re-earn the click. Open on the promise or the most curiosity-loaded moment: cold open, mid-action, or the claim itself. Never open on branding, intros, channel housekeeping, or "hey guys." Restate the title's promise visually or verbally within those seconds so the click and the content lock together, then foreshadow the best moment: "wait until you see what happens at the end."
2. **Reengagement.** Immediately after the hook, give reasons to stay: open a second loop, tease a later payoff, start delivering value fast. Viewers decide again at 30-60 seconds whether this video keeps its promise; that's where the first big audience leak happens, so the first minute carries real content, not setup.
3. **The setup.** Build toward the payoff. Don't rush to the climax; keep teasing it while delivering on the way. Weave a secondary thread through the middle so there's always an open loop pulling forward.
4. **The climax.** The payoff the packaging promised. Deliver it fully and don't undercut it; an underdelivered climax is the broken promise that kills a channel's trust over time.
5. **The bonus beat.** Something extra after the payoff: a gem, a twist, a last laugh, one more piece of value the viewer didn't expect. Done consistently, it trains the audience to watch to the very end because they know something's still coming, exactly like a post-credits scene.
6. **The wrap-up.** Close the loop from the hook, fast. Outros are where viewers leave, so keep it seconds long, no "that's it for today," and point straight at the next video with a verbal call to action plus an end screen to a genuinely related video.

The working shorthand for the whole arc: hook them, tell a story, evoke an emotion, pay it off at the end, keep them watching all the way through. Script with that sentence taped to the monitor.

Scripting rules that hold everywhere:

- **Message first, content second** for anything that sells; content first for pure entertainment. Never sacrifice the message for a joke.
- **Don't be married to anything.** Cut scenes that don't serve the message even when they're funny. Emotional detachment in the edit is a competitive advantage.
- **Write for the ear.** Short sentences, concrete words, the audience's own vocabulary. Read the script aloud; anything that trips the tongue trips the viewer.
- **Mark the interrupts in the script.** Plan a pattern change roughly every 20-40 seconds of talking-head content (less necessary when the footage itself keeps changing): an angle switch, a cut-away, a prop, a joke, a location change. Planned interrupts beat rescue edits.

## Phase 5: pacing and the edit

Retention is the metric the algorithm feeds on. The edit's job is to keep the retention line flat.

- **Respect the micro-commitment.** A click is a viewer saying no to everything else on the internet. Every second re-earns it; the working standard is hours of creative effort per second of viewer attention.
- **Maintain the value proposition throughout.** Length is not the enemy; value-less stretches are. A four-minute ad can hold over 90% of viewers to the end when every step keeps delivering. Cut everything that doesn't move the promise forward: long intros, repeated points, breathing room that's really just dead air.
- **Pattern interrupts at the drift points.** Watch the cut as a viewer and mark every moment attention drifts, then interrupt: a camera angle change or zoom (the strongest single tool), a cut to b-roll or something unexpected, a sound effect, a beat of humor, new information teasing what's ahead, a hand clap, a voice change. The best channels in the world do this relentlessly; study any top creator's video with a stopwatch and count the visual changes per minute.
- **Watch the trigger phrases.** "Don't miss what's at the end" reengages; anything that smells like wrapping up ("so yeah", "that's basically it", "before we finish") triggers an exodus. Never signal the end before the end.
- **Sound design is retention.** Music shifts, risers into reveals, silence before the payoff. Sound cues re-grab wandering attention as effectively as cuts do.
- **Speed is a tool.** A slight global speed-up, even a few percent, tightens pacing without viewers noticing, and trimming every pause tightens it more.

## Phase 6: read the data, then adjust

Publish, then study. The cycle is try, fail, analyze, adjust, and it never stops.

- **The two retention metrics:** average view duration (AVD, time watched) and average view percentage (AVP, share of the video watched). They diverge naturally (playback speed, video length), so grow them together: packaging gets the views, but AVD and AVP build the audience and the recommendations.
- **CTR in context:** rising impressions with falling CTR is usually good news, because it means the video is being shown beyond the core audience. Judge CTR against the traffic source, not as one global number, and remember browse, search, and suggested each have different baselines.
- **The 30% check:** find where the first 30% of viewers left, usually inside the first 30-60 seconds. That's a hook or expectation problem: the open didn't match what the packaging promised. Fix the open in the next video.
- **The 50% check:** find the point where half the audience is gone, watch that exact section, and name what triggered the leak. The whole game is moving that point later, video after video.
- **Study the bumps, not just the dips.** Rises and flat stretches in the retention graph are the moments that worked. Identify what happened there (a reveal, a joke, a format change, a chapter viewers jump to) and deliberately do more of it. Also check the graph against video chapters: chapters people skip to are topics the audience wants standalone videos about.
- **Diagnose by shape.** A cliff in the first 15 seconds (the hockey-stick graph) is a packaging-to-content disconnect. A slow steady bleed is pacing. A sharp drop at a specific timestamp is a specific mistake, so go watch it. A drop at the end is a bloated outro.
- **Test cadence:** two thumbnails per video via Test & Compare, two different hooks per short. Publish variants in different styles, find the top three formats by response, ride the winners while continuing to test new ones. Never change three things at once on a struggling channel, because then nothing is learned.
- **Review rhythm:** deep-dive the analytics on a fixed schedule, start of month and mid-month, checking each video at first hour, first day, first week, and 30 days. Compare only against the channel's own history and its niche, because every audience behaves differently.
- **Feed it back to Phase 0.** The channel's own outliers go into the same dataset as competitors'. What over-performed against the channel median is the audience voting; make the next video for the voters.
- **Volume is the tuition.** Nobody's first videos are good, and the first hundred are practice. The realistic path is shipping a lot, analyzing every one, and improving a little each time. The analysis habit, more than talent, is what separates channels that compound from channels that stall.

## Running this skill

When the user is starting or growing a channel: begin at Phase 0, build the competitor set, run the scrape, and deliver the outlier list, the pattern analysis, and ten validated video ideas ranked by demand evidence. When the user brings a video idea: run Phase 1 and be willing to kill the idea if no packaging survives, then work the phases in order, producing a title shortlist with the conversation test applied, two or three thumbnail concepts described in enough detail to shoot, a script outline mapped to the six-beat arc with pattern interrupts marked, and an edit checklist. When the user brings an underperforming video: start at Phase 6, diagnose from the retention shape and the 30%/50% points, and trace the problem back to the phase that caused it. When the user asks for ideas: pull from the outlier list and comment mining, never from generic brainstorming alone.