Key takeaways
- A script is not a word count, it is a retention plan written down before it is expensive to change. The cheapest edit in the world is deleting a paragraph that has not been filmed yet.
- Since 24 August 2026 a YouTube view is counted from the first frame, and the older, stricter definition is now called an engaged view. Packaging buys the play; the script decides whether the play becomes the number that pays.
- The structure worth copying from the leaked MrBeast production handbook is not the budget — it is that every block of minutes has a named job, and the job changes roughly every three minutes.
- Curiosity is not infinite. Loewenstein's information-gap account says it peaks when the viewer already knows a little, which is why an open loop over an unexplained situation does nothing at all.
- Most sagging middles are an information problem, not a pacing problem. Faster cuts on a segment that reveals nothing make the drop steeper, because the viewer reaches the conclusion sooner.
- Script length has an arithmetic answer: at a conversational 150 words a minute, a ten-minute video is roughly 1,300 to 1,500 spoken words, and almost every first draft is well over it.
There is a specific kind of video that makes this worth writing about. The thumbnail is good. The title is good. Click-through is above anything the channel has done before, impressions keep arriving for two days, and the video still goes nowhere. The creator concludes the algorithm did not like it, makes another thumbnail, and repeats the experiment.
What actually happened is visible in the retention curve, and it is nearly always the same shape: a healthy opening, a soft slope, then a section in the middle where the line bends down and never recovers. That section was written — or, more often, not written. It was improvised in front of a camera, kept because it was filmed, and defended in the edit because cutting it meant admitting the shoot had produced eleven usable minutes out of nineteen.
Scripting is the cheapest intervention available to a channel that already knows how to package. It costs an afternoon, no equipment and no new skill, and it operates on the half of the funnel that thumbnails cannot touch. This is how to do it in a way that survives contact with a camera.
What a script is actually for
The word carries baggage. Say "script" to most creators and they picture a word-for-word document read off a teleprompter in the flat cadence that makes a video feel like a pharmaceutical advert. That is one implementation, often the wrong one.
A script is the only place where the shape of a video is cheap to change. Once a thing is filmed, every structural decision is negotiated against sunk effort: the segment that does not earn its four minutes is defended because it took three hours to shoot, the anecdote that goes nowhere survives because it went well on the day. On paper, the same segment costs one keystroke to remove. The second function follows from the first — a script lets you state the promise and then check whether you kept it. You cannot see, halfway through a shoot, that minute seven has quietly stopped being about the thing the thumbnail advertised. You can see it in an outline in about four seconds.
The promise is the spec
Write the title and design the thumbnail before you write the script, not after. This inverts the usual order and it is the single highest-leverage change in this entire post.
The packaging is a contract. It names a specific thing a viewer will get, and the script is the document that delivers it. Write the script first and you end up reverse-engineering a promise out of whatever you happened to film, which is how channels arrive at vague titles — the video was about six things, so the title had to be about none of them. Write the packaging first and every scripting decision has a test attached: does this paragraph move the viewer towards the thing the thumbnail showed, or away from it?
It also kills bad ideas early. A packaging idea no script can pay off is a bad packaging idea, and you find that out on paper rather than after the upload. If the thumbnail shows an outcome you never reach on camera, the retention curve punishes it precisely, which our post on the psychology of clickbait thumbnails takes apart in more detail.
Why the middle is worth more than it used to be
Until August 2026, a long-form view was only counted after a viewer had stayed for roughly half a minute, so the public counter had a retention filter quietly built into it. From 24 August 2026 that changed: views are counted from the first frame across every format, and the old definition survives under a new name as the engaged view — a play where someone stayed past the opening seconds or did something deliberate like liking, commenting or sharing.
The practical consequence is a widening gap between two numbers that used to be one number. The public count now reflects how well the packaging worked; the engaged count, which is what monetisation and recommendation continue to run on, reflects what the video did after the click. Two videos with identical thumbnails and identical click-through can now produce visibly different engaged-view counts, and the difference is written in the script. None of that changes the underlying craft. It changes who can see the craft failing, and how quickly.
What the MrBeast handbook actually says about structure
In 2024 a 36-page internal document from MrBeast Productions, titled "How to Succeed in MrBeast Production", leaked and was widely reported — Tubefilter's summary is the most sober of the write-ups. Most of the coverage focused on the employment sections. The structurally interesting part was ignored.
The document treats a video as a sequence of blocks with assigned jobs. The first minute is named as the most important one, because that is where viewers leave in the largest numbers; the guidance is to match the expectation the packaging set and front-load information. Minutes one to three are the transition from hype to execution, using what the document calls "crazy progression" — if the video is somebody surviving weeks in a forest, the first three minutes cover several days rather than lingering on day one. Around the three-minute mark it calls for a wow moment. Minutes three to six and six onward carry their own defined responsibilities in turn.
One number is worth sitting with. It describes losing 21 million viewers in the first minute of a video with roughly 60 million clicks, and calls that better than average — a production operation with resources no reader of this post has, reporting a first-minute loss of about a third and calling it a good day.
What transfers to a channel with no budget is not the spectacle. It is the discipline of assigning a job to each block of minutes before writing a word of dialogue. A channel making twelve-minute explainers can run the same exercise on an index card: minute zero to one, pay off the thumbnail; one to three, establish the stakes and the first concrete result; three to six, the hard part done properly; six to ten, the complication nobody expects; ten to end, resolution and the next thing to watch. The blocks are arbitrary until you write them down, at which point they become a test each segment either passes or fails.
Why middles sag, in three causes
Open any retention graph with a mid-video collapse and the cause is one of three things. They look identical on the curve and need completely different fixes.
| Cause | What it looks like on the page | The fix |
|---|---|---|
| No new information | A segment that restates, recaps or elaborates something already established | Delete it. Not shorten — delete |
| No change in stakes | The situation at minute nine is the situation at minute five, with more words | Introduce a complication, a constraint or a failure |
| No structural signal | Correct, dense content delivered in one undifferentiated block | Break it: a chapter, a location change, a question asked out loud |
The third one has a research basis worth knowing about, because it explains why the advice to "cut faster" sometimes works and sometimes makes things measurably worse. Annie Lang's limited-capacity model of mediated message processing, developed across two decades of communication research, treats a viewer as a processor with a fixed pool of attention that is allocated partly automatically. Structural features of the message itself — cuts, edits, changes in pitch, movement, anything that constitutes an abrupt change — elicit an orienting response and pull resources back to the screen. Content that is motivationally relevant does the same.
That is the mechanism behind every "add a cut every four seconds" tip, and it is also the reason the tip fails when applied to the wrong problem. An orienting response buys you attention for the next moment, not interest in the next minute. Point it at a segment that has nothing to reveal and the extra attention is spent confirming there is nothing to stay for. Faster cutting over hollow content does not flatten the drop; it sharpens it, because you have helped the viewer reach the conclusion sooner. Structure is the amplifier, not the signal.
Open loops, and the part everyone gets wrong
The standard advice is to open a loop early and close it late. The advice is sound and the usual execution is not, because the mechanism has a precondition that nobody mentions.
George Loewenstein's 1994 review in Psychological Bulletin framed curiosity as an information gap: a felt deprivation that appears when attention lands on a hole in what you know. The part that matters for scripting is the shape of the curve. Curiosity is at its strongest when someone already has a moderate amount of knowledge about a subject, and it collapses at both ends — when they know nothing, and when they know everything. A gap you cannot perceive is not a gap. It is just absence.
Which means an open loop only works after the viewer has enough context to feel what is missing. "You won't believe what happened next" placed forty seconds into a video, before anyone knows who is involved or what is at stake, creates no tension at all — there is no structure in the viewer's head for the missing piece to be missing from. The same line at minute six, after the situation is established and the stakes are clear, is doing real work.
The loop ledger
Write every loop you open on one line of a list, with the timestamp where it opens and the timestamp where it closes. Two rules. Every loop closes inside the video, not in a future upload and not in the comments. No more than two are open at once — a third makes the viewer stop tracking them, at which point none of them are holding anything. If a loop has no closing timestamp, it is not a loop, it is a broken promise with better phrasing.
Re-hooks: what to put where the line bends
A re-hook is a deliberate re-statement of why the rest of the video is worth staying for, placed where you expect people to leave. In a scripted document they are planned, not improvised, and a handful of shapes cover almost every case.
| Shape | What it does | Where it belongs |
|---|---|---|
| Forward reference | Names a specific thing coming later, with enough detail to be wanted | Just before a necessary but dull segment |
| Complication | Something fails, or a constraint appears that was not there before | The midpoint, where stakes usually flatten |
| Result delivered early | Gives away an outcome to buy attention for how it happened | Anywhere the process is longer than the payoff |
| Direct address | Names the viewer's likely objection and answers it | After a claim that strains credibility |
| Hard reset | Location, format or register changes entirely | Long videos, roughly every five to eight minutes |
Cadence matters more than count. A re-hook every ninety seconds reads as a nervous video that cannot commit to a point. One at each block boundary is usually right, plus one wherever your own back catalogue says viewers leave — which is knowable rather than guessable, and is what the audience retention graph is for.
Chapters are the cheapest structural signal available, and they do double duty: they break the undifferentiated block, and they give the video multiple entry points in search. The requirements are strict enough that many creators believe they have chapters when they do not, which our post on chapters and key moments covers properly.
Writing for the ear, with arithmetic
Scripts are written on a page and consumed as speech, and the two have different tolerances. Written prose can hold a subordinate clause three levels deep. Speech cannot, because the listener has no way to look back at the start of the sentence.
The useful constraint is length. The National Center for Voice and Speech puts average conversational English in the United States at around 150 words a minute, and most creators land somewhere between 130 and 170 depending on energy and editing style. That gives you a direct conversion from runtime to word budget.
| Target runtime | Spoken words at ~150 wpm | What that is, in practice |
|---|---|---|
| 60 seconds | ~150 | A Short, or one segment of a long-form video |
| 8 minutes | ~1,200 | A tight single-topic explainer |
| 12 minutes | ~1,800 | A tutorial with a worked example |
| 20 minutes | ~3,000 | A documentary-style or essay format |
The numbers expose padding immediately. If your twelve-minute tutorial has 2,900 words of narration, you are not making a twelve-minute video; you are making a twenty-minute video that will be rushed on camera and still overrun, and the overrun lands in the middle where it does the most damage. Most first drafts come in at roughly double budget, which is fine — the draft is for finding the video, the cut is for making it. On which length actually suits a given format, our post on how long a YouTube video should be goes through the trade-offs.
Three habits fix most read-aloud problems. Read the draft out loud and mark every place you run out of breath — those sentences need splitting, not better delivery. Cut the first sentence of every paragraph and check whether anything was lost; usually the paragraph was warming up. Replace abstract nouns with the concrete thing they stand for, because a listener cannot re-read "optimisation" to work out which one you meant.
Word-for-word, beats, or both
The right level of scripting depends on the format and on how you perform, not on which method a creator you admire uses. The honest version of this decision is a table.
| Method | Works for | Fails at |
|---|---|---|
| Full word-for-word | Voiceover, essays, faceless formats, tightly argued explainers | On-camera delivery without practice — it reads as recitation |
| Beat sheet | Talking head, reviews, anything with a personality carrying it | Precision. Numbers, names and legal wording drift |
| Hybrid | Most channels: scripted hook, ending and transitions, bullets elsewhere | Nothing much, which is why it is the default recommendation |
| Unscripted | Interviews, live, vlogs, reaction formats | Structure. It must be imposed in the edit instead, which is slower |
The hybrid deserves its position. The segments that benefit most from exact wording are the ones where a fumble is expensive: the opening, the transitions where improvisers wander, and the ending. Everything between those can be bullets if bullets suit you, provided each bullet states the point rather than the topic. "Cost" is a topic. "This costs about four times what people expect, and here is the invoice" is a point.
A spine for formats that are not stories
Narrative structure gets recommended indiscriminately, including for videos that have no story in them. It still transfers, but the reason is worth stating precisely rather than mystically.
Green and Brock's 2000 experiments in the Journal of Personality and Social Psychology introduced narrative transportation: readers absorbed into a story held more story-consistent beliefs and, in the second experiment, noticed fewer false notes than less-transported readers. Absorption reduces counterarguing. For a creator the relevant consequence is not persuasion but attention — a viewer tracking an unresolved situation is not simultaneously deciding whether to leave.
So the spine to impose on a non-story video is a question with a delayed answer, not a plot. A tutorial becomes: here is the outcome, here is what makes it harder than it looks, here is the sequence, here is the failure mode, here is the result. A review becomes: here is the claim, here is the test, here is where it breaks. List videos resist this hardest, which is why they sag so reliably — the format tells the viewer the items are interchangeable, so leaving after item four costs nothing. The fix is ordering that escalates and a stated reason to reach the end.
The attention-span number you should stop repeating
Someone will tell you that humans now have an attention span of eight seconds, less than a goldfish, and that videos must be built accordingly. That figure came from a 2015 Microsoft Canada report, spread through several national newspapers, and when the BBC investigated it in 2017 the trail led to a firm called Statistic Brain that could not produce a credible source. The Microsoft document itself does not contain the claim. There is no goldfish.
What does exist is Gloria Mark's fieldwork at UC Irvine, which measured how long people stay on a single screen before switching. In research published in 2004 the average was around two and a half minutes; by the mid-2010s her measurements put it at about 47 seconds, with a median around 40. She is careful, in interviews such as the American Psychological Association's Speaking of Psychology episode, that this measures switching behaviour in a distracting environment rather than a biological ceiling on human attention.
The distinction matters because the two readings give opposite advice. The myth says nothing holds anyone beyond a few seconds, so chop everything. The research says people switch away from screens that give them a reason to, on devices engineered to offer alternatives — and it sits alongside the ordinary observation that the same population watches three-hour podcasts. Your video is not competing against a biological limit. It is competing against the next thumbnail, and it loses at the moments where it stops giving anyone a reason to stay.
The ending is a script problem too
Most endings are an apology for the video finishing. The script should treat the last ninety seconds as its own block with two jobs: land the argument the video actually made, and hand the viewer a specific next thing.
Specific is the operative word. "Check out my other videos" performs about as well as saying nothing, because it asks the viewer to do the selection work. Naming one video, and why it follows from this one, converts differently — and it is the part of the ending you can write in advance. The surfaces that carry it are covered in our post on end screens and cards; the script's job is to give them a sentence worth attaching to.
One thing not to script: a long outro. The end screen period is already dead retention time in every curve, and a monologue over it drags average view duration down without buying anything.
Shorts invert the whole structure. Leaving costs one thumb movement and the thing is often watched twice, so the shape that works is payoff-first, context-second, with an ending fast enough to seed a rewatch — inside roughly 150 words for a sixty-second Short.
How to tell whether the script worked
Scripts cannot be A/B tested. Thumbnails can, which is why creators end up with better instincts about images than about structure. What you have instead is a slower comparison across your own uploads.
Three measurements, in order of usefulness. First, the retention curve at the block boundaries you defined — if the line drops where you planned a re-hook, you know which sixty seconds to rewrite. Second, average view duration against your own median for that runtime, the only benchmark that means anything; cross-channel averages for this are invented. Third, the ratio of views to engaged views, which since August 2026 reads how many plays survived the opening.
Change one structural thing at a time and give it three or four uploads. The sample is small and the variance is large, and the standard mistake is to rewrite the whole approach after one video that underperformed for reasons unrelated to the script. Our post on the first thirty seconds covers how to test an opening specifically, which is the one block where the sample arrives fast enough to learn from quickly.
Seven ways a script goes wrong
- Writing before the packaging exists. The video ends up about six things, and the title has to be vague enough to cover all of them.
- Front-loading credentials. Nobody has decided you are worth listening to yet, so the qualifications you lead with are spent on an audience that has not stayed.
- Recapping. A viewer who is still there does not need a summary of the last four minutes, and one who left is not watching your recap.
- Loops that never close. Answering a question in a future video teaches your audience that your loops do not pay, and they stop responding to them.
- Writing to a runtime. Stretching a seven-minute idea to hit an arbitrary length shows on the curve within ninety seconds of where the padding starts.
- Scripting the words but not the visuals. Narration-only scripts produce the video where a talking head reads a well-written essay over nothing.
- Never reading it aloud. Every stumble in the shoot was already in the document, and would have cost nothing to fix there.
A workflow that fits in one afternoon
Pull it together into something repeatable. Decide the packaging first, with the promise stated in one sentence. Block out the runtime with a job assigned to each block, and mark where the re-hooks sit. Draft at roughly twice your word budget without editing, then cut back to budget by removing whole segments before trimming sentences. Write the hook, the transitions and the last ninety seconds word for word; leave the rest at whatever level of detail you can perform. Read it aloud with a timer. Then shoot it — and afterwards put the retention curve next to the document and mark where the line disagreed with the plan.
That last step is the one that compounds. A script is a hypothesis about where attention will hold, and the curve is the result of the experiment. Most creators never put the two side by side, which is why the same structural mistake survives fifty uploads.
The half of the job this does not do
Everything here operates after the click. It is worth being blunt about the limit: a perfectly structured video that nobody opens is worth precisely nothing, and no amount of scripting discipline recovers a packaging failure. The two halves are not interchangeable and they are not in competition — impressions are won by the thumbnail and title, and engaged views are won by what follows.
The reason to fix both is that the damage compounds in one direction. Good packaging on a video that does not deliver trains an audience to distrust your thumbnails, and that shows up as falling click-through on a channel whose images have not changed. The promise has to be decided once, at the start, and then kept by both assets.
That is the part of the workflow worth making fast. If designing the thumbnail is a two-hour job, it happens after the edit, and the promise gets reverse-engineered out of the footage — the exact failure this post opened with. Thumblore exists to collapse that step to a few minutes, so the packaging can be settled before the script is written and the script can be written to deliver it. Get the order right and the rest of this is ordinary craft: decide what you are promising, work out where a viewer would otherwise leave, and put something there worth staying for.