Video Script

Video Script

The body is everything after the hook has done its work. It decides whether the viewer is still there at minute twelve, and whether the one ask at the end lands on somebody who has already decided.

This skill does not write the opening. The first 30 seconds, the hook variants and the drop-off rating are video-hooks, and its measured finding stands: the hook governs whether people stay, never how many arrive. Read it first, take the open from it, and come back for everything after 0:30.

The body's structure is storytelling. The five moves, the literalness ladder, one open loop and the motif that returns changed. The finding it exists to prevent is the script that is accurate all the way through and holds nobody: read the beats in pairs and say what changed between each one, because "another point about the same thing" is a list, and a list is watched at the speed of the scrub bar.

Voice is voice-dna. Do not restate it, do not soften it, and run humanizer over the finished script before it is read on camera. Count the first-person pronouns first: the target is zero.

Read what the feed rewards this month, before drafting

social-engagement derives it from competitor posts it collects itself, scored against each account's own median so a 22,000-like account and a 21-like account contribute the same evidence: social-engagement, references/what-works.md. It carries its measurement date and a staleness ladder, because a feed re-ranks every month and a file does not. Over 30 days old, collect the feed again before trusting the page.

It says what the feed rewards, never what is good. The bans in @heyramzi/lint and the voice in voice-dna outrank every number on that page.

Decide the subject on a number before writing a beat

idea-mining chooses what the piece is about. This is the measurement that says whether it is worth a recording day, and it runs before the hook.

One read, and it is free.

• The SERP, and it decides alone. yt-dlp on the exact phrase a buyer types, then a second pass for the counts, because --flat-playlist returns NA for them:

  yt-dlp --flat-playlist --skip-download "ytsearch20:<phrase>" --print "%(id)s" > ids.txt
  yt-dlp --skip-download --no-warnings -a ids.txt --print "%(id)s|%(view_count)s|%(title)s"
  

  Read the view counts, not the titles. A first page of 2-to-5-minute clips under 150 views each is an unserved query, and the only condition that justifies entering a keyword a bigger channel already holds.

• Demand is no longer measurable, so do not fake it. yt-dlp returns no search volume or competition score. Decide on the SERP, and say in one line that volume was not measured.

Worked example, 28 Aug 2026, EC52. clickup pricing returned 3,541 monthly searches at competition 21.3, a middling row, and the SERP returned 11, 7, 27, 118 and 16 views on its top five. The second read is what made it a video. The data cost policy is programmatic-seo's and it binds here too, so do not spend a paid keyword call on a subject idea-mining has not already picked.

A fallback that invents its own data is worse than an outage. The app's keyword report answered four different seeds with an identical estimated volume of 140 and seven templated ideas while looking exactly like a measurement, and a title was chosen off it on 26 Aug 2026. Any keyword row repeating across unrelated seeds is a fallback, not a reading.

The runtime rules live in code

Where an app generates a script, the rules it enforces live in code. That file is the runtime copy and this page is the evidence behind it. Change one and change the other in the same session, because a rule that only exists here never reaches a generated script, and one that only exists there has nothing to defend it.

The three-run control is the spine of this skill

Runtime, beat budget and ask placement were each measured across three runs of the channel, and the numbers are what this skill enforces. Read references/three-run-control.md before writing the beat sheet, or when a script has come out long and the budget has to be argued with.

The ask rule, where two sources of evidence disagree

The generic retention literature says never save the only CTA for the end, because only 16% of viewers reach the final 10%. The three-run control says the two winners held every outbound ask to the last 2% and the loser asked at 11% and lost by 3 to 5x. Both are right, because they are counting different asks.

• An in-platform ask (subscribe, like, comment) costs the viewer nothing and never breaks the teach. Place these where the retention data says: one around the one-minute mark after the first thing of value has landed, one more later, each paired with something on screen, because embedded beats spoken-only by a wide margin.
• An outbound ask (buy this, book a call, go to this page) spends the viewer's attention and their trust at once. There is exactly one, in the last twenty seconds. A second does not add a second conversion, it converts the video into an ad in the viewer's memory.

If the video exists to sell something, that is an argument for making the teach complete, not for asking earlier. The winning run teaches for nineteen minutes and asks once with 21 seconds left.

Two further sources corroborate the split and one of them bounds it: Saraev's four-hour course with its single closing pitch, the 627-transcript measurement that puts the median last outbound ask at 92% of runtime and 77% of videos on one ask or none, and the reason that corpus cannot settle the rule either way. All three, with the corrections they force on the mid-roll advice above, are in references/three-run-control.md.

A native embed is not an ask, and it is the only free mention in the script

Third category, sitting under both rows above. A native embed is a real thing named inside the teach because the argument arrived at it: the tool that solved the step being demonstrated, the artefact the viewer would need to do this themselves, the community the worked example came from. It sends the viewer nowhere and interrupts nothing, so it does not spend what an outbound ask spends and the one-ask rule has never counted it.

The distinction is whether the mention breaks the frame. A viewer twelve minutes into a teach has stopped noticing they are watching a video, and a break to a different scene, shirt or tone wakes them up, at which point they remember they have work to do. A sentence that continues the argument does not. So set up the block until the thing being named is the answer to the question it just raised, and let it stay in the teach. Native embeds may appear early and may repeat; the outbound ask may not, and is still exactly one, still in the last twenty seconds.

Where a sponsor read cannot be made native, it goes after the average view duration for the channel, never before it, so the interruption lands on viewers who were leaving anyway. This is the only placement rule in this skill that is not derived from the three-run control; it is carried in from the Kallaway corpus and it is not control-tested. attention-mechanics records the collision and why the one-ask rule survives it intact.

Two hits before the first exit, and the eyes-closed test

Two audits, both run on the draft rather than on the analytics.

Two usable things, early. One non-obvious thing the viewer can go and change lands inside the first block, a second before the average view duration. One hit and they may stay; two and they stay, because the video has proved it holds more. A definition is not a hit: the test is whether they can act on it today, rung 3 on storytelling's literalness ladder.

The eyes-closed test settles pacing, and nothing else does. Play the cut with the picture off and listen. Bored means the sentences are running long and the edit is not chopping; unable to keep up means it is chopped past comprehension and needs breathing room put back. The page has no tempo, so reading it finds neither. Run it on the first assembly, and the note goes to the editor as a note about the audio.

Reset your tolerance before you judge the cut. After watching a take twenty times you are the least qualified viewer it has: scroll a feed for five minutes, then watch it once through. What reads as slow on the second pass is slow, and the correction usually lands in the first thirty seconds. Single source, no control test: Mino, Feb 2026, video-hooks references/hook-teardown.md, who reports cutting roughly half a finished video this way.

One debatable question, written on purpose. A block that leaves a genuinely contested question open produces comment threads that argue with each other, and argument is the engagement platforms pay for. Not a bait question addressed to the audience, which generate-social bans and X demotes: a real fork in the subject that the video declines to close, said once, in the teach.

The seam between two blocks is where they leave

A body loses people where a block finishes, because that is the moment nothing is unresolved and the viewer remembers they have work to do. So every seam carries a re-hook: the last beat of the block closes it and opens the next question in the same breath, the way a relay baton changes hands while both runners are moving.

The target is no stretch past 300 seconds with nothing reopened, measured 29 Aug 2026 over 339 punctuated transcripts: the niche's median gap between loop openers is 230 seconds and its p10 is 85, so 300 is tighter than 90% of what ships. It replaces the 60-to-90-second cadence this page carried for one day, which was one creator's unmeasured number.

The test is to delete the transition words: a fact or a question still standing means it was a re-hook, and a sentence that disappears was throat-clearing. Those phrases (here's the thing, it turns out, the truth is) are banned in slop-words.js, which 75% of the niche never uses.

The seam is the re-hook beat of storytelling, references/addiction-loop.md, which also carries what the block owes before it: stakes with a clock in them, and a big question specific enough for the viewer to guess wrong at. A body with no point where the viewer can guess wrong is a correct list, and a bullet outline is the cheapest place to fix that.

Nothing here is judged by eye

python3 .claude/skills/video-script/scripts/story_metrics.py <script.json|take.txt> --duration <s> --grade
python3 .claude/skills/video-script/scripts/teardown.py <url>          # a reference video, same axes

story_metrics.py --grade returns seven rules in two tiers: FAIL is a banned construction or a number outside what the niche tolerates, WARN is inside the niche and short of the house target. Every threshold prints its source, three being percentiles of 627 measured videos and four doctrine. 59% of that corpus passes clean and each rule fires on 5 to 20%, which is what makes a fail mean something, and it refuses to grade the sentence-shaped rules on an unpunctuated transcript rather than inventing a number. What each rule is, what teardown.py prints, and the norms behind both: references/measuring.md.

Then run the six story locks over the finished bullets, one pass each: term branding, embedded truths, thought narration, negative frames, loop openers, contrast words, in video-hooks, references/story-locks.md. Three land hardest on a body. Thought narration: say the thought the viewer is having, in their words, from customer-voice rather than guessed. Embedded truths: hedge the provenance, never the instruction. Contrast words: split the sentences carrying the main points and turn the viewer inside them, which is the zigzag at sentence scale.

The script names the overlays, and only its own

The CTA shelf at /design carries one clip per sellable product plus the on-camera asks. Which of them an edit needs is decided by the ask this script writes, so the script returns them by file name in overlays and the concept page shows those files alone. Handing the editor the whole shelf hands them somebody else's product, and the right clip is then missed among fifteen wrong ones.

One clip for the outbound ask, the one that names what the ask names. A subscribe, like or comment clip only where a block actually makes that ask on camera. A name that is not on the shelf is dropped rather than shown, because a tile that leads nowhere is a download the editor chases.

What a series buys

One breakout, plus episodes that hold the people already captured. Six ran 17,874, 11,425, 4,144, 5,209, 960, 2,258, and episode five landed below the channel median. Front-load accordingly: episode one is the reach event, everything after is conversion work on an audience you already have.

The script is one of five surfaces

A recorded video is cut into five surfaces and the long-form script is only the first. What each surface takes from it and what it must not repeat: references/five-surfaces.md, read once the script is written and the video goes into production.

Writing it

The bullet rule, the beat budget, the structure, the teach block that is 53% of the script, the camera-language line and the register: references/writing-the-script.md.

The same file holds the argument video, the format where there is no build to teach: the dated chain whose every link ends on the lack that forces the next, the motif that comes back as the answer, the caveat spent on the number in place, the fenced prediction and the close that gives permission before the ask. Read it whenever the body is a claim about the world rather than a screen with a click in it. Single unmeasured reference, so it is candidate craft sitting under everything the three-run control decided.

Before it ships

• Beat count inside 158-185, and the runtime that implies stated at the top.
• No bullet past 14 words unless the whole line is a quotation.
• Exactly one outbound ask, inside the final twenty seconds. Native embeds are not counted, and any non-native sponsor read sits after the channel's average view duration.
• Two things the viewer can act on today, the first inside the opening block and the second before the average view duration.
• The eyes-closed pass was run on the first assembly and the pacing note went to the editor.
• story_metrics.py --grade run, with zero fails and every warn either fixed or answered in a line.
• One point where the viewer forms a prediction and one where it breaks, with the clues already on screen. The six locks were run, and every hedge left in is provenance rather than instruction.
• The failure told in full, not summarised.
• Every number traced to something measured, with the artefact named for the edit.
• No arithmetic performed on camera.
• One portable idea, named in the mechanism and pointed at as each build lands, and every build block opening on its own one-sentence claim at a third to a half of the block.
• An argument video ran the essay checks: every history link ends on the lack that forces the next, any prediction block is fenced out loud, and the close gives permission before the ask.
• At least one abstraction drawn on screen rather than described.
• Every term defined the first time it is said, in objects the viewer already owns.
• One thing shown failing inside each build block, uncut.
• The dating answer present wherever the subject is a tool that ships weekly.
• One chapter per block, named for a state rather than a feature. The corpus: 78% publish chapters, median 7 and 3.9 per ten minutes, titled in 4 words, the first ending at 4.7% of runtime.
• humanizer run over the whole thing.

Self-Healing

This skill appends new failure modes to references/learned-patterns.md after each run, newest last. A finding that turns out to be stale gets corrected here in the same session, along with the generator's own copy of the rule if it enforces one.