Video Script Writing

---
name: video-script-writing
description: >-
  Write and revise video scripts that hold viewer attention to the end and
  drive the desired action, grounded in retention and persuasion research
  (information-gap curiosity, Zeigarnik open loops, hook-concreteness studies,
  narrative transportation, speech-rate and scene-cut experiments). Use when
  the user asks for a video script, YouTube script, or explainer script; a
  founder, brand-story, or talking-head video; a landing-page, marketing,
  demo, or testimonial video; a video ad or VSL; or a short-form script
  (Shorts, Reels, TikTok). Also use when improving an existing script's hook,
  pacing, retention, or CTA, or when asked whether a script will hold
  attention. Covers hook tests, loop architecture, spoken-register drafting,
  production specs (cut rhythm, captions, words-per-minute bands, placement),
  and per-purpose playbooks with benchmarks.
license: MIT
---

# Video Script Writing

Retention-first script writing for any video purpose. The script is the
product; production quality is secondary. A rough, real, well-structured
script beats a polished script with weak architecture.

Distilled 2026-08-26 from the Learn Anything founder-video session plus a
research pass over hook studies, curiosity psychology, narrative
transportation, speech-rate experiments, and eye-tracking editing research.
Full citations: `references/research-digest.md`.

## Pair with

- A voice skill for the speaker's prose voice (e.g. write-like-eric) when
  writing as a specific person; the script must survive a table read.
- marketing-copy for offer construction and persuasion strategy upstream.
- eric-video (or the production pipeline du jour) when the script moves to
  rendering; this skill ends at the script plus its production spec.

## Step 1 — Intake (ask before drafting)

- Purpose (founder story, conversion, explainer, short-form, demo)?
  Selects the playbook.
- Where does it live (by a form, YouTube, feed)? Placement, CTA shape,
  captions.
- Who watches, and what is their ONE unresolved doubt? The video exists
  to kill that doubt; without one, recommend against a video.
- Length budget? Sets loop spacing and wpm math.
- Whose voice, and what evidence/receipts do they own? Numbers come from
  the speaker's own verifiable material.
- The single action after watching? One CTA per script, no exceptions.

## Step 2 — Choose the playbook

Read the matching section of `references/purpose-playbooks.md` BEFORE
drafting:

| Purpose | Length | Primary metric |
| --- | --- | --- |
| Founder / brand story | 2 to 3 min | Trust + waitlist/signup |
| Landing-page conversion | 60 to 90 s | Conversion lift vs no video |
| YouTube explainer / education | 5 to 10 min | Retention curve shape |
| Short-form (Shorts/Reels/TikTok) | 15 to 60 s | Completion + rewatch |
| Demo / testimonial | 60 to 120 s | Doubt killed, next step taken |

## Step 3 — Architecture: draw the loop map first

Before writing prose, produce a loop map table (loop, where it opens, where
it closes). Rules, all research-backed (see digest):

- One macro loop: opens in the first 15 seconds, closes in the final 20
  percent. Closing it early collapses the reason to keep watching.
- Mid loops at section boundaries; micro-satisfactions (a payoff, a number, a
  reveal, a laugh) at least every 10 to 15 seconds of script.
- At most 2 loops open concurrently; more overloads working memory.
- Every loop CLOSES by the end. Unresolved loops read as betrayal and damage
  the next video too (Zeigarnik works on grudges).
- Place a deliberate re-engagement beat (stakes escalation, specificity
  injection, perspective shift, social proof) at 20 to 30 percent of word
  count: the known flat zone where momentum fades before commitment sets in.
- A single curiosity gap holds pull for 3 to 4 minutes max; renew or resolve.
- Cognitive load ceiling: 4 to 5 new concepts per minute; break any section
  past ~300 words; prefer adversative transitions ("but", "the catch") over
  additive ones ("also", "next").
- Peak-end: put the best moment near the end and end warm with one specific
  next action, never on a summary or a subscribe beg (viewers skip closings,
  thanks, and subscribe prompts; this is measured).

## Step 4 — The hook (first 15 seconds)

Draft the opening last, against the actual promise. It must pass all three
tests (349-video hit-vs-flop study):

1. Reason test: a concrete reason to keep watching (claim, result, question,
   or stake) lands within ~15 seconds. "Today we talk about X" is a topic,
   not a reason.
2. Number test: one real, checkable figure early. Hits carry numbers 63% vs
   52% in flops. Pull it from the speaker's own verified material.
3. Promise test: the opening engages the exact promise the title/thumbnail
   or page headline made. Never open adjacent to it.

Failure modes to strike on sight: context dumps (background before a reason
to care; a quarter of all flops), greetings ("hey everyone, welcome back"),
and promise-only openers ("stick around and I'll show you" asks for credit;
give instead).

GENRE EXCEPTION - founder / invitation / landing-page videos (2026-08-27,
Learn Anything founder video): when the video is a PERSONAL ADDRESS (founder
story, invitation, landing-page embed - the viewer pressed play on a PERSON,
not a topic), a warm greeting opener is CORRECT, not a flop. Eric explicitly
commissioned: "Hey everybody, I'm Eric. I thought I'd take a little bit of
time to tell you a little bit of my convictions and motivations for running
this retreat." The greeting + one-line intent IS the promised content here,
so it passes the promise test; the greeting flop-pattern data comes from
topic-promise YouTube videos where the title/thumbnail claim must be engaged
in ~15s. Corollary (same session): closers in this genre carry NO
honesty-flag hedges ("And honestly, ...") - "if I have to write honestly,
that means I wasn't being honest" - close warm and conversational.
Dictation rules apply: splice Eric's dictated opener/closer phrasing
near-verbatim.

Mechanics: make the stake second person ("you") where honest; size the gap
right (enough context to make the gap visible, not so much it self-closes;
curiosity follows an inverted-U); explanatory gaps ("I know what, not why")
pull 2 to 3x longer than surface gaps ("what happens next").

Null finding, keep in mind: within the normal delivery band, speaking pace
and speed-to-point do NOT separate hits from flops. Concreteness does. Do
not fix a weak hook by talking faster.

## Step 5 — Draft in spoken register

- Write for the mouth: short speakable sentences, contractions, no
  semicolon-prose, no written-register curriculum-speak. Read every line
  aloud mentally; if the speaker wouldn't say it to a friend, rewrite.
- Target the purpose's wpm band (playbooks table). Persuasion is curvilinear
  in rate: moderately brisk wins, slow reads as low-competence, frantic
  overloads. Trim dead air in the edit instead of rushing delivery.
- Concrete receipts over adjectives: real numbers, named events, owned
  evidence ("I did X, here is the artifact").
- Vary sentence lengths; land thesis sentences as their own short sentence.
- If the user rejects drafts as "not my voice", stop generating variants and
  ask them to dictate or table-read; their spoken phrasing is authoritative
  and gets spliced near-verbatim.

## Step 6 — Attach the production spec

Every script ships with a spec block (the retention levers live in the edit
as much as the words):

- Visual change (cut, b-roll, overlay, angle, micro-zoom) at least every 4
  to 7 seconds; every extra second of an unchanged scene measurably drops
  attentional focus. No static 5-second windows anywhere in the plan.
- Put each key line immediately AFTER a visual change: attention peaks about
  two-thirds of a second after a cut. Hold important on-screen elements at
  least 1 second, centered, uncluttered (visual complexity depresses focus).
- Seamless jump cuts (dead-air removal) for fluency; use sparing overlap
  only to re-engage, and never chase cut-speed for its own sake (sustained
  attention falls at high transition frequency).
- Burn in captions, always: most mobile viewers watch muted and captioned
  videos finish far more often. The script must work captions-only.
- Click-to-play with a face-forward custom thumbnail; sound-on autoplay
  never. Lazy-load the player so the page CTA renders first.
- Landing/conversion surfaces: video sits beside the CTA, the CTA stays
  visible on a 375px screen when the video ends, video stays ungated, end
  screen points at the single next action.

## Step 7 — Pre-ship verification

- [ ] Muted test: the whole argument survives captions alone.
- [ ] Math: per-section word counts over timestamps land inside the wpm band.
- [ ] Loop map complete; every loop closed; never more than 2 concurrent.
- [ ] Hook passes reason, number, and promise tests within 15 seconds.
- [ ] Exactly one CTA; the end screen or closing line points at it.
- [ ] No planned static window over 5 seconds; key lines follow cuts.
- [ ] Claims audit: promise only what the offerer controls; no invented
      urgency; numbers verified against the speaker's own material.
- [ ] Table read done by the actual speaker; flagged lines rewritten in
      their dictated phrasing.

## Step 8 — Measure after launch

Strong completion targets by length: under 5 min, 65 to 75 percent; 5 to 10
min, 50 to 60; 10 to 15 min, 40 to 50. Embedded conversion video: 60 to 75
percent completion is strong. Cliffs cluster in the first 15 seconds (hook)
and at the 25 to 35 percent mark (missing re-engagement beat); when a cliff
appears, fix that moment, not the whole script.

## Honesty rules

- Video is conversion-neutral ON AVERAGE across landing pages (large
  dataset). It pays when it kills the named doubt from Step 1; otherwise it
  is weight on the page. A/B test rather than assume lift.
- No curiosity bait the video does not pay off inside the video. The
  resolution lives in the watch, never in a caption, comment, or "part 2".
- Do not over-cut. Diminishing returns are real; processing takes time;
  scenes under one second cannot encode.
- Engagement is not conversion. Judge the script by the action it produces.

## Weaving source material into an existing script (revision protocol)

Distilled 2026-08-29 from the Learn Anything founder-video session (mining
the Daniel Fischer intake-call transcript for story material into a
runtime-capped script).

1. Mine with one read-only reporter, assess in the main thread. When a
   large transcript/source feeds an existing script, save the source to a
   file and dispatch ONE explore subagent with a three-part brief:
   (a) the headline candidate stories VERBATIM, (b) all material mapping
   onto the script's existing beats, (c) anything CONTRADICTING current
   script claims. The main thread holds the constraint set (the script's
   one doubt-killing job, runtime budget, guardrails, retention rules)
   and does the assessment; the subagent only extracts.
2. Assess every candidate against four axes BEFORE editing: structure
   fit, runtime (additions to an over-budget script require trades),
   guardrails (provenance, claims audit), retention (concreteness,
   caption legibility). A candidate that fails an axis is parked WITH the
   reason logged, not silently dropped.
3. Paired cut (show-don't-tell swap): when an inserted story
   INSTANTIATES an abstract claim already in the script ("each one took
   real effort and time"), cut the abstract claim in the same edit — the
   story now carries it. Net runtime stays flat; assertion becomes
   evidence.
4. Keep a trim ladder in the script notes: an ordered list of
   pre-approved cuts, weakest first (e.g. field names → lever/crutch
   couplet → never the struggle), so future runtime overruns trade down
   deterministically instead of re-deliberating.
5. Provenance fidelity beats rhetorical punch: NEVER transplant a
   verbatim quote from one story arc into another (a "five more times"
   quote from the backprop arc must not land in the Gaussian-processes
   arc even though it is the stronger line). Record each woven story's
   provenance (source, date) in the script's guardrails block.
6. Earnings history as transformation, never figures: in
   founder-marketing scripts the speaker's own outcome is admissible as
   history ("that tripled my grad student stipend") but dollar figures
   ($112K) and job titles are omitted — concrete money numbers read as
   implied outcome promises. The speaker's own hedge on the source call
   ("that's not a lot, right?") is the tell the number will land wrong.
7. Auto-transcript numbers need a table-read flag: near-verbatim quotes
   from an auto-generated transcript carry garbled counts and names;
   adjacent story arcs can hold different counts ("five more times" vs
   "two, three times"). Flag every numeric or named claim sourced from
   auto-transcript for speaker confirmation at the table read.