Write or audit short-form and long-form video scripts using a fixed packaging → outline → intro → body → outro skeleton with a 2-1-3-4 body order, value loops, rehooks, native CTAs, and a live middle-first / hook-last process for 60-90s videos. Use for TikTok / Reels / Shorts / vertical YouTube / long-form YouTube / demo videos / launch videos / pitch openers — wherever a script needs to keep viewers hooked end to end.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Write or audit short-form and long-form video scripts using a fixed packaging → outline → intro → body → outro skeleton with a 2-1-3-4 body order, value loops, rehooks, native CTAs, and a live middle-first / hook-last process for 60-90s videos. Use for TikTok / Reels / Shorts / vertical YouTube / long-form YouTube / demo videos / launch videos / pitch openers — wherever a script needs to keep viewers hooked end to end.
Use this skill to script videos that beat expectations. Every viral script is the same machine: the title sets an expectation; every following beat must push reality slightly above it. That single gap drives retention, likes, shares, and rewatches. Cleverness is downstream.
There are two complementary layers. The 5-step skeleton (packaging → outline → intro → body → outro) governs every script regardless of length. The 6-phase live process (source intake → middle first → hook last → iterate aloud → cut for time → lock) governs how a script actually gets written when you're staring at a blank page. The skeleton is the diagram; the process is the work under your hands.
This is a procedure skill at domain altitude: two ordered length-specific workflows (long-form and short-form) run over that skeleton and live process, with an audit mode for existing drafts.
Activation
Draft a long-form YouTube script (8–25 min) with a fixed structure.
Draft a short-form video script (60–90s) where the middle-first process matters more than the structural template.
Audit an existing draft: identify which step of the skeleton failed, then rewrite that step.
Convert a brain dump or research seed into a script with a single defensible plot line.
Iterate a script that "feels off" — diagnose by reading aloud, not by re-reading silently.
Do not use this skill for hook craft alone (defer to hook_craft_short_form), for narrative shape and curiosity-loop architecture alone (defer to story_driven_content_craft), or for the upstream "what to write about" decision (defer to nonfiction_writing_from_lived_conviction and content_strategy_beyond_blogging). This skill is the script-structure layer between those upstream decisions and the published video. Full ownership map in Routing.
Judgment
Core Principles
Reality must beat expectations. Every line either raises reality vs. the title's promise, holds it, or shrinks it. Holding is losing.
Order is non-negotiable. Packaging → outline → intro → body → outro. The order is a forcing function — it stops you from polishing prose before you have a unique point.
Title locked, thumbnail loose. Drafting before the title is locked makes click confirmation impossible. Thumbnail perfectionism kills script momentum.
Outline before prose. The outline step exists to gut-check uniqueness. Beautiful sentences cannot fix unoriginal points.
Pick one plot line. Especially in short-form. "You can't answer everything in 60 seconds." Combination plot lines are the most common short-form failure.
Write the middle first. Capture inspired lines during research. The hook has to promise what the middle delivers — write the promise after the deliverable exists.
Iterate the hook 8–10 times. Always with reading aloud. Voice-first iteration is the actual mechanism.
2-1-3-4, never 1-2-3-4. Put your second-best body point first, your best second, then descend. The ascending pattern is the retention mechanism, not any single point.
Read aloud after every change. Silent revision = bad rhythm. The page is a proxy; the ear is the truth.
Cut for time, even when it hurts. Over 90 seconds in short-form? Cut from expansions, never from the hook or the take. Conviction beats coverage.
Why 2-1-3-4 (the ordering rule)
Human brains read patterns. Two ascending points train the viewer that value is increasing — they stay for point three. Lead with your best and the slope flips downward; the viewer subconsciously expects diminishing returns and bounces. Music albums put the single in slot 3 or 4 for the same reason.
For multi-section content beyond video — threads, decks, blog posts, demo flows — apply the same pattern: second-strongest opens, strongest pays off, then descend.
BuildOS Voice Translation
The Kallaway register skews toward shock and creator-economy hype. BuildOS positions anti-AI, anti-feed, anti-hype. The mechanics hold; the register changes.
The contrarian take in the 5-part intro is BuildOS's natural strength. Default contrasts: AI-as-autopilot vs. AI-as-thinking-environment; manufactured virality vs. interest media; productivity-app churn vs. stable thinking environment; task-list theater vs. structured thinking output.
The verdict ending in short-form should land as conviction-from-experience, not bombast. "This is the only way I've found to keep my projects from rotting" beats "This will change everything."
The "win-win frame" for takes maps cleanly to BuildOS's anti-zero-sum positioning — most BuildOS POV pieces are arguments that someone's losing assumption is wrong but their underlying need is real.
The native CTA anchors to BuildOS resources naturally: the brain-dump flow as the solve to "I have ideas piling up but nothing structured," the daily brief as the solve to "I context-switch ten times a day," the project pages as the solve to "I lose context across weeks."
No manufactured urgency or scarcity in the outro. "Free chocolate" is allowed; "act now or miss out" is not.
The 2-1-3-4 reordering still applies, but the metric of "novelty + shock + uniqueness" should bias toward novelty + lived proof + non-obvious framing rather than shock alone.
Word-by-word audit must catch BuildOS-specific jargon: "ontology," "context engineering," "agent-native," "ambient compute" are all teenager-test fails. Translate to working vocabulary or kill.
Procedure
Layer 1 — The 5-Step Skeleton (governs every script, scales by length)
STEP 0 — PSYCHOLOGY (governing principle, not a step)
Reality must beat the expectation set by the title.
Score every beat: did this raise reality, hold it, or shrink it?
STEP 1 — PACKAGING
Idea: one-sentence pain point or curiosity rabbit hole for the ideal viewer
Title: locked; sets the expectation contract
Thumbnail: loose; placeholder is fine
STEP 2 — OUTLINE
Bulleted points before any prose
Each bullet: what / why / how
Uniqueness check: is this novel vs common takes in the niche?
If no, return to research. Do NOT proceed.
STEP 3 — INTRO (5-part hook formula)
(1) Immediate context + click confirmation
(2) Common belief / conventional take
(3) Contrarian take that breaks the common belief
(4) Plan / ordered structure of what's coming
(5) Proof / credibility (4 and 5 swappable)
STEP 4 — BODY (2-1-3-4 ordering)
Rank points by raw novelty + shock + uniqueness
Reorder: 2nd-best, best, 3rd, 4th, 5th…
For each point, run the Value Loop:
- Context (what — simplest possible statement)
- Application (how — multiple concrete examples)
- Framing (why — zoomed-out meaning)
Between points: insert a rehook (mini stake-raising transition)
Rhythm: alternate zoom-in (tactic) and zoom-out (big picture)
STEP 5 — OUTRO (Fortune Cookie)
Recap the points
Restate the pain that was solved
End on a high emotional note
Goal: trigger likes, shares, comments, rewatches
BONUS — NATIVE CTA EMBEDS
Tag 1–2 body points where a free resource = natural pain solve
Embed CTA as continuation of the point, never as "subscribe break"
Test: read the script with the CTA removed — does it still feel complete?
Layer 2 — The 6-Phase Live Process (how scripts actually get written)
This is what you do when you sit down with a blank doc. It is the active form of the skeleton — it shows what each step looks like when you're working under your own hands. Especially load-bearing for short-form (60–90s), where compression makes structural shortcuts dangerous.
PHASE 1 — SOURCE INTAKE (~6 min for short-form, longer for long-form)
Open the seed (tweet, link, screenshot, brain dump)
Skim once, react out loud — capture confusion and "wait, what?" moments
Validate signal: source engagement, prior viral hits in the same theme
Pull 1–2 secondary sources at increasing simplification levels
Identify 4–6 possible plot lines
LOCK ONE plot line. Verbally name what's getting cut.
PHASE 2 — MIDDLE FIRST (~8 min)
Capture inspired phrases verbatim as they emerge
Draft the explanation: "here's what's crazy" + concrete examples
Add the founder's POV / take after the explanation
Don't worry about transitions yet
Don't worry about the hook yet
PHASE 3 — HOOK LAST (~13 min)
Open with subject + simplification
Anchor with prior credibility if available
Iterate 8–10 times, reading aloud each pass
Default to the SHORTEST version that earns its job
Test: "if a viewer hears this hook, do they know exactly what they'll learn next?"
PHASE 4 — ITERATE AND PACE (~15 min)
Read full draft aloud start to end
Mark lines that don't dance — too long, too similar, too jargon
Replace abstract entities with concrete renamings (the simulation → AI Netflix)
Add bridge phrases ("picture this", "well now they're back")
Cut adjacent ideas creeping back in
PHASE 5 — ENDING (~6 min)
State the verdict with conviction
Optional: tease one of the cut angles in passing
End on a prediction line, not a CTA
PHASE 6 — FINAL READ + CUT FOR TIME
Time the read aloud (~150 wpm)
Cut expansions first, never the hook or the take
Lock and record
Workflow: Long-Form Script (8–25 min YouTube)
Lock the title. Sets the expectation contract.
Outline the points. Bullets, not prose. Each bullet: what / why / how.
Run uniqueness check. Compare against existing top videos in the niche. If common knowledge, return to research. Do not draft.
Rank points and reorder 2-1-3-4. Justify in writing why point 2 beats point 1 on novelty + shock + uniqueness.
Write the 5-part intro. All five blocks labeled before stitching to prose. Most commonly skipped: common belief.
Write the body. Each point gets a Value Loop (context / application / framing). Insert a rehook between every two points. Reject generic transitions ("Next, let's talk about…").
Write the outro (Fortune Cookie). Recap → pain solved → high emotional note. Never end with "anyway, that's it."
Tag 1–2 native CTA embeds. Anchored to body points where the free resource is the natural solve. Test: read the script without the CTA — does it still feel complete?
Defer hook polish. → hook_craft_short_form. The 5-part intro is the structural shape; the hook-craft skill handles the slot grammar / archetype / 4-mistake diagnostic on the opening lines specifically.
Run the dopamine ladder check. → story_driven_content_craft. Articulate the implanted question at the 5-second mark and at the body opener. Confirm the validation on each loop close is non-obvious.
For short-form, structure compresses but the live process becomes more important — there is no room for unforced errors.
Validate signal. What's the proof this idea will resonate? (Source engagement, prior similar wins, friend tag, gut + receipts.) If no signal, push back before drafting.
Pull simplification layers. Source → secondary → simpler restatement. Rename abstract entities ("the simulation" → "AI Netflix") until a non-expert teenager could track it.
List 4–6 possible plot lines. Force the founder to name them. Pick one for retellability and emotional payoff. Verbally name what's getting cut.
Capture inspired phrases verbatim during research. That's where the actual lines live.
Write the middle first. What is it + concrete examples + the founder's take. Don't transition; don't hook.
Layer the founder's POV on top of the news. "What's happening + my take of the business / strategy angle" is the moat. Conviction reads as authority.
Build the win-win frame for the take (or whatever framing the take needs). 3 bullets compressed to 3 sentences. Structure makes the take feel earned, not loud.
Surface the obvious objection and dismiss it. One sentence. Inoculates against pushback before the verdict.
Write and iterate the hook 8–10 times. Always reading aloud. Default to the SHORTEST that earns its job — trust the next line to do the explaining.
End on a verdict, not a CTA. "This is going to be massive and could completely change Hollywood." Conviction. Cut "I think" and "maybe."
Read aloud, time against 90 seconds. Cut from expansions, never from the hook or the take.
Tease one cut angle near the end (optional). "There's so much more here including their 10 original shows." Single line. Often gets cut from final.
Lock and record.
For the hook itself, defer to hook_craft_short_form — the slot grammar (Subject / Action / Objective / Contrast / Proof / Time), the 6-archetype catalog, and the 4-mistake diagnostic apply.
Routing
Ownership map. The Procedure sequences and defers via → markers; this table assigns. One concept, one owner — this skill routes deep hook/story/upstream ownership out.
Body — each point with Value Loop and rehook into next
Outro — recap + pain solved + high note
Native CTA tags — which body points anchor which CTAs (1–2 max for long-form, 1 max for short-form)
Hook variants — ONLY when no locked hook was supplied: 5–8 iterations with the chosen one and read-aloud notes. If a locked hook IS supplied, skip variants entirely, carry the locked hook through unchanged, and note "locked hook — variants deferred to hook_craft_short_form if wanted."
Cadence map — sentence-length pattern with bridge phrases marked
Word audit — flagged jargon / abstract entities replaced
Read time — vs. target window (e.g., 1:42 / 1:30 = cut 12s from expansions)
Conviction ending — verdict line, hedges stripped
Expectation-vs-reality scorecard — for each beat: raise / hold / shrink
For audit-mode runs, replace the bundle with a diagnostic report: which step of the skeleton failed (packaging / outline / intro / body / outro / CTA), which phase of the live process broke (signal / simplification / plot line / middle / hook / pacing / cut), and the specific rewrite plan.
Policy
No drafting before the title is locked. No exceptions.
No prose before the outline. No exceptions.
No script that fails the uniqueness check. Beautiful sentences cannot fix unoriginal points.
No 1-2-3-4 ordering. Every script reordered to 2-1-3-4-5-… Justify the swap explicitly.
No body point missing the Value Loop. All three components (context / application / framing) must be present.
No generic transitions. "Next, let's talk about…" = bounce trigger. Rehook or don't transition.
No outro that ends with "anyway, that's it." Recap → pain solved → high emotional note, every time.
No CTA that breaks the spell. Native embeds only. Test by removing the CTA — script must still read complete.
No combination plot lines in short-form. One plot line per video. Reject 2 even if both are great.
No abstract entity names. Rename until a teenager could track it.
No silent revision. Read aloud after every meaningful edit.
No hedging in the ending. Strip "I think," "maybe," "probably," "kind of." Conviction beats coverage.
No script over 90s in short-form. Cut from expansions; never from the hook or the take.
No hook locked before the middle is real. The hook promises what the middle delivers. Promise without delivery = bait.
No founder vocabulary unedited. Audit every domain word as if a teenager is listening.
Knowledge
The craft schemas below are PRIMARY — distilled from the two Kallaway videos in Provenance.
The 5-part intro, line by line
Block
Job
Watch out
Immediate context
Confirm the click. Name the topic from the title plainly.
Don't open with throat-clearing or a setup question.
Common belief
State the conventional take — the viewer's existing baseline.
Most commonly skipped block. Without it, the contrarian has nothing to push against.
Contrarian take
Contradict the common belief. This is where reality starts to beat expectations.
Must be a real opposing claim, not "but actually it's nuanced."
Plan
Ordered structure of what's coming (steps, list, framework name).
Names a framework? Even better. Anchor it.
Proof
Credibility — why trust this contrarian approach.
Keep tight. Proof is a stamp, not a résumé.
The Value Loop, per body point
POINT N
Context — what it is, simplest possible statement
Application — how to do it, multiple concrete examples
Framing — why it matters, how it slots into the bigger story
REHOOK — "that was important, but if you don't pair it with this next
one, the whole thing collapses"
POINT N+1
The "dance" — sentence cadence
Pattern that recurs in winning short-form scripts:
LONG (15–20 words: subject + simplification)
SHORT (2–3 words: "it's crazy", "picture this")
LONG (the explanation)
SHORT–MEDIUM ("well now they're back")
LONG (the payoff scenario)
SHORT ("how insane is that")
LONG (the take)
SHORT–MEDIUM (verdict)
Vary length deliberately. Punchy-punchy-punchy reads monotone. Long-long-long reads sleepy. Read aloud after every revision.
Word-by-word audit checklist
Any word the founder wouldn't say in casual conversation? → swap.
Any abstract entity name? → replace with concrete renaming (the simulation → AI Netflix).
Any sentence over 25 words? → break or cut.
4+ similar-length sentences in a row? → insert a short beat.
Any "I think / maybe / probably" in the ending? → cut for conviction.
Read time over 90 seconds (short-form)? → cut from expansions.
Cross-Platform Compression
The 5-step skeleton scales by length. Same skeleton, different proportions:
Platform
Intro
Body
Outro
Hook iter
Long-form YouTube (8–25 min)
Full 5-part
4–7 points, full Value Loops, 1–2 native CTAs
Full Fortune Cookie
5–8 passes
Mid-form (3–8 min)
Compressed 5-part (skip common belief if obvious)
3–4 points, tighter Value Loops, 1 native CTA
Tight Fortune Cookie
5 passes
Short-form (60–90s)
First 3s = click confirm + raise; common belief implied
2–3 quick value beats, 1 stake, 1 take
Single high-note button (verdict)
8–10 passes
Threads / LinkedIn carousel
Post 1 = 5-part compressed
Posts 2-N = Value Loops with rehooks
Final post = Fortune Cookie + native CTA
3–5 passes
Landing page
Hero = click confirm + contrarian take
Features = 2-1-3-4 ordering
FAQ + final CTA
3–5 passes
Pitch deck
Title + problem + insight
Traction → 2-1-3-4 value drivers
Vision slide
5 passes
Demo video
First 5s = "watch this run"
Each step = Value Loop
"And here's what just happened"
5 passes
Brain dump → blog post
Lede = 5-part compressed
2-1-3-4 sub-headed sections
Final paragraph kicker
3 passes
Examples
Worked Example
Condensed gold-standard bundle for the eval Task 1 fixture (locked hook, 60–90s vertical, brain-dump demo footage, overwhelmed-founder audience); input in evals.md. Hook is locked upstream, so hook variants are skipped and the 5-part intro compresses to the first-3s click confirm. Match this shape.
Title (working caption): "Stop organizing your notes. Watch this instead." — locked before drafting.
Thumbnail/cover brief (loose): mid-transformation board frame, overlay "Mess in, plan out."
Signal check: passed — a rough 20s clip of the same transformation pulled 4× usual saves.
Plot line (locked): organizing-at-capture is the tax that kills every productivity app; skip it and structure shows up on its own — demonstrated live. Cut angles, named: daily brief, calendar sync, AI-agents-changing-founder-work riff (teased nowhere; kept for separate videos).
Ranked points (2-1-3-4): Best = the live ramble→plan transformation (lived proof on screen). Second-best = why every app failed you: they make you organize while you capture. Third = the capture-first take. Reorder justification: the why-apps-fail point opens because it re-agitates the hook's pain and trains an ascending-value slope into the demo; leading with the demo would flip the slope downward after 20 seconds.
Body (Value Loops + rehooks):
Point A — organizing at capture is the tax (talking-head)
Context: "Every productivity app makes you organize while you capture."
Application: "Pick the project. Pick the tag. Pick the folder. Mid-thought."
Framing: "That tax is why you quit every app in three weeks."
Rehook: "But if organizing at capture is the problem — watch what happens when you skip it entirely."
Point B — the live transformation (screen recording)
Context: "This is one raw voice ramble. Zero structure." (transcript on screen)
Framing: "Forty seconds ago this was a mess in my head. Now it's a plan I can start."
Rehook: "And here's the part nobody tells you."
Point C — the take (talking-head)
Context: "You never had a discipline problem."
Application: "You had a capture problem. Dump everything, every morning, one place."
Framing: "Structure isn't something you impose. It shows up once everything's finally in one spot."
Objection + dismissal: "Sounds too simple? That's exactly why I almost didn't ship it."
Conviction ending (verdict, hedges stripped): "This is the only way I've found to stop my projects from rotting across six apps." (ends on the finished board — pays the hook's promise exactly)
Native CTA tag: none in the spoken script; pinned-comment link to the brain-dump flow, anchored to Point C's pain. Removal test: script reads complete without it — pass.
Cadence map: A = LONG/SHORT-staccato/MEDIUM · rehook SHORT · B = MEDIUM/SHORT-SHORT-SHORT/LONG · rehook SHORT · C = SHORT/MEDIUM/LONG · objection SHORT · verdict MEDIUM. Read-aloud pass: no 4-in-a-row same-length runs.
Word audit: "context engineering" / "ontology" never used; "AI structures it" renamed to "tasks pull themselves out"; every line passes the would-I-say-it-out-loud test.
Read time: ~190 words incl. hook ≈ 0:76 at 150 wpm — inside 90s; no cuts needed (expansion to cut first if it ran long: the three-staccato line in Point A).
Expectation-vs-reality scorecard: hook = raise · A-context = hold · A-framing = raise · rehook 1 = raise · B-application = raise (the payoff) · B-framing = hold · rehook 2 = raise · C = raise (reframe) · objection = hold · verdict = raise.
Provenance
Distilled from two Kallaway videos (PRIMARY — creator source):
How To Write A Killer Script That Keeps Viewers Hooked — the 5-step skeleton (packaging / outline / intro / body / outro), the 5-part intro hook formula, the 2-1-3-4 body ordering, the Value Loop (context / application / framing), the rehook between points, the Fortune Cookie outro, and the native CTA embed pattern. Plus the governing expectations-vs-reality principle.
Watch Me Write a VIRAL Tiktok Script (Full Breakdown) — the 6-phase live process (source intake → middle first → hook last → iterate aloud → cut for time → lock), the "pick one plot line" rule, the abstract-to-concrete renaming discipline, the sentence "dance" cadence, the 8–10× hook iteration loop, the win-win frame for takes, the surface-the-objection move, and the verdict ending.
Sibling ownership (what each paired skill owns) is declared in Routing.
