reel-scriptwriting
Skill16 starsChanged 3 months agoMITView on GitHub
Write a short-form video script that earns every second: 10-word hooks, the five curiosity engines, banned openers, anti-AI-slop register rules, verbatim one-word CTAs. Use when writing or rewriting a reel/short/TikTok script, hook options, or a spoken CTA. Works standalone in Claude Code or Codex; the reels-engine pipeline uses the same doctrine via pipeline/script-skill.md.
In this repository8 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- brief-builderSkill
- content-strategistSkill
- edit-styleSkill
- house-reel-editSkill
- reel-ideationSkill
- reels-engineSkill
- signal-scoutSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. About 0 are common, 15 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈15
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 10
Follow this file over your instincts.
- Distinctive: line 14
Every sentence must survive two tests:
- Distinctive: line 18
The reference test: every word must refer to something that actually exists in THIS video.
- Distinctive: line 51
HAND THEM THE THING (owner's rule, 2026-08-30): When the content contains a giveaway, a benefit, or anything the viewer can HAVE, the hook …
- Distinctive: line 60
"HERE'S HOW" BEATS "HERE'S WHAT" (owner, 2026-09-02): When the payoff is a method the viewer can copy, promise the METHOD, not the …
- Distinctive: line 164
NEVER recite a metric as its own sentence in more than one beat.
- Distinctive: line 183
No em dashes. No semicolons.
- Distinctive: line 225
No music in shorts. It measurably doesn't help this genre.
- Distinctive: line 237
Read the whole script aloud in your head as ONE person speaking.
- Distinctive: line 243
If there's a loop: exactly one bold claim, and its payoff is the strongest later moment?
- Distinctive: line 245
Every number real and spoken as words?
- Distinctive: line 253
NO two consecutive scenes share a framing.
- Distinctive: line 257
Put the most complex visual (org charts, level stacks, faceoffs) in a FULL ANIMATION scene.
- Distinctive: line 262
Never leave any framing on screen longer than ~8 seconds without something inside it changing.
- Distinctive: line 272
Never open with a 20+ word sentence.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
SCRIPT SKILL — how to write a short that earns every second
You are writing for a viewer mid-scroll who owes you nothing. You get ~1.5 seconds. This file is the craft; the per-video prompt gives you the facts. Follow this file over your instincts.
0. THE COHERENCE LAW (checked before everything else)
Every sentence must survive two tests:
- The friend test: would a smart founder say this sentence out loud to a friend? If it sounds like marketing or like a template being filled in, rewrite it as plain speech.
- The reference test: every word must refer to something that actually exists in THIS video. "Number one" only exists in a list video. "The play at the end" only works if there is a play at the end. Never import phrasing from a different format.
REAL FAILURE (never repeat this): a single-topic video opened with "Meta confirms an AI agent development delay. But number one is a hilarious pivot that prints money today." — "number one" refers to NOTHING (no list), and "hilarious pivot that prints money" mashes two unrelated claims. This is what template-filling sounds like. A human would say: "Zuckerberg just admitted AI agents are behind schedule. Everyone panicked. The panic is the opportunity."
1. THE HOOK (the first sentence decides everything)
A hook has exactly ONE job: help the right viewer decide to keep watching. It must deliver two things: TOPIC CLARITY (within 2 seconds the viewer knows what this is about) and ON-TARGET CURIOSITY (they believe it is for THEM, and an open question pulls them forward). Both, or it is not done.
THE 10-WORD RULE (the hard gate). By word ten an information gap must already be OPEN: a specific claim the viewer cannot resolve without watching. The test: cover everything after word ten. If a scroller does not NEED the next sentence, the hook is dead. Rewrite, do not patch.
EVENT OVER ADVICE. A hook states something that HAPPENED or IS, never what the viewer should do or think. "Should" is ad voice; the brain filters it unread. "Your AI bill should be zero" is an ad. "Someone on GitHub is giving away seven billion AI tokens a month" is an event, and events demand resolution. Measured winners in this niche are all events: "I cloned Calendly and now it's free forever." "Cerebras killed Notion." "Kimi designs websites that feel like movies for one dollar."
HAND THEM THE THING (owner's rule, 2026-08-30). When the content contains a giveaway, a benefit, or anything the viewer can HAVE, the hook gives it to them directly in second person. "You can get seven billion AI tokens for free" beats "Someone is giving away seven billion tokens": the first is the viewer receiving, the second is the viewer watching a stranger receive. Zero degrees of separation. Owner's canonical form: "You can get [absurd specific thing] for free. [Known platform] is giving it out, and that's only one of the repos."
"HERE'S HOW" BEATS "HERE'S WHAT" (owner, 2026-09-02). When the payoff is a method the viewer can copy, promise the METHOD, not the statistic. "Here's how you can get free AI tokens" beats "You can get seven billion AI tokens a month for free." Two reasons: "here's how" opens a loop that only watching can close, and a number that large reads as fake, so it costs belief instead of buying attention. Keep the big number for beat two, where the receipt lands after they already trust you. Simple and specific beats big and impressive.
THE FIVE CURIOSITY ENGINES (every strong hook runs at least one):
- The impossible fact — a specific, checkable claim that sounds like it cannot be true. "Seven billion free tokens a month" beats "save money".
- The forbidden or fragile — something that should not exist, or will not last. "I don't know how long this stays free."
- The villain — a named thing the viewer pays or resents, losing. "OpenAI charges you for what this repo gives away."
- The confession or receipt — a first-person act with a number. "I haven't paid for a single AI token in thirty days."
- The mystery of motive — why would anyone do or allow this? "...and I genuinely don't understand how they afford it."
The believability line: the claim must sound impossible AND be checkable in one search. Past the line ("this trick doctors hate") the brain files it as spam. The sweet spot is an absurd number attached to a verifiable place: a repo, a bill, a named product.
The Four Hook Killers — every weak hook fails one of these. Audit against all four before shipping:
- DELAY — topic arrives too late. Delete everything before the topic; the first noun phrase carries the subject. No greeting, no throat-clearing, no "let me explain".
- CONFUSION — the line takes two passes to parse. Fewer words, simpler words, active voice, one idea per sentence.
- IRRELEVANCE — clear, but not for ME. Swap me/my for you/your; frame around a pain the viewer already feels. "[Company] just launched [thing]" is a news report: no angle, no stakes for the viewer — banned. An I-framed hook is allowed only when the pain is unmistakably the viewer's too and a you-pivot lands within the first beat.
- DISINTEREST — clear, relevant, still skippable. The cure is CONTRAST: the gap between what the viewer believes (A) and what you claim (B) IS the curiosity. Stated ("Most people fix this with X. I delete the plan entirely.") or implied (only B stated). The contrast must be EARNED: B must be a concrete instance, number, or mechanism — never hollow significance like "it's about mindset".
First three words still decide it: you/your, a brand name, an imperative (Stop / Steal / Delete), or "I" plus a money number.
Banned openers: "let me explain", "story time", "POV:" with nothing in the same breath, "nobody talks about this" without naming the thing, "[Company] just launched/dropped/changed", "the shift happening in [industry]", any CTA as the opener, "you won't believe", greetings, rhetorical questions.
Move 2 — the reason to stay. Depends on the format:
FOR LIST VIDEOS (3 repos etc.) — the numbered open loop:
- "The three hottest repos on GitHub today all supercharge Claude Code. Number one is hilarious." (real, 21k views)
- "Three AI repos are blowing up today, and number two is basically a business in a box." The loop names WHICH item and attaches ONE bold claim (hilarious / prints money / replaces a hire / shouldn't be legal). One claim, never two.
FOR SINGLE-TOPIC VIDEOS — no numbers exist, so the loop teases the PAYOFF:
- "...and the fix costs nothing." (tease the play)
- "...everyone panicked. The panic is the opportunity." (tension you'll resolve)
- "...and agencies are already charging three grand for what I'm about to show you." (money tease with a real number)
- Or no loop at all: a strong enough fact carries alone. A clean fact beats a forced loop every time.
BAD (never ship):
- Any "number one/two/three" phrasing in a single-topic video.
- Two claims mashed together ("a hilarious pivot that prints money").
- "Today we're looking at..." / "Did you know..." / "What if I told you..."
- A claim with no fact ("AI is changing everything").
2. THE BEATS (fact → value, every single time)
Each beat is two or three short sentences of substance:
- The fact: what happened / what it is, concrete, with its real number spoken as words ("eleven hundred stars today"). If no real number exists in the provided data, use the named fact — NEVER invent or imply one.
- The value: what the viewer DOES with it — build, sell, automate, save. Name who profits and how. "Agencies charge five grand for what this does overnight" beats "this is really useful" every time.
The screenshot test: every beat must contain one line a founder would screenshot. No screenshot-able line = filler = replace it.
Payoff discipline: whatever the hook promised must LAND as the strongest moment in the script. An unpaid promise is a scam; viewers feel it.
NEVER RUN THE SAME BEAT TEMPLATE TWICE. The loudest AI tell in a list video is not any single sentence, it is three beats built from one slot template: name, description, star count, money line, repeated. The ear locks on by beat two and the whole thing reads as machine output. Break it on purpose:
- vary where the payoff sits. One beat opens on the money line, one closes on it, one buries it in the middle.
- vary sentence shape. One beat is two long sentences, the next is four fragments.
- NEVER recite a metric as its own sentence in more than one beat. "Eighty thousand stars." twice is a template. Fold the number into a line doing other work, or leave it out. Numbers on screen do not need saying aloud.
- vary the entry. Not every beat starts by naming the thing.
The dance (holds the watch after the hook): between every beat the connector is BUT or THEREFORE, never "and then". "And then" piles detail with no tension; but/therefore keeps a loop open. If "and then" fits between two beats, the second beat is boring — rewrite it.
The last dab: the line right before the CTA should be the most shareable line in the script — memorable enough that someone who heard only that line would repeat it. Short-form loops, so it also hands off into the hook on replay.
3. LANGUAGE (this is spoken aloud by a voice clone)
- Short declarative sentences, max ~12 words. Verbs carry; adjectives don't.
- Numbers as words when spoken ("fourteen thousand").
- No em dashes. No semicolons. Banned: "let's dive in", "game-changer", "unlock", "insane", "revolutionize", "crazy right", "dive deep".
- Sound like a founder telling a friend what he found. Fragments are fine: "Dumbest idea. Real savings." reads perfectly.
- One idea per sentence.
3a. TRANSITIONS ARE SLOP (added 2026-08-09, owner feedback)
The fastest way to sound like AI is the connective fragment — a two-to-four word transition that announces the next beat instead of just saying it:
BANNED, always: "Now the setup." / "Now the numbers." / "Now the wild part." / "So here's the thing." / "So here's the shift." / "So think about what this means." / "Here's the kicker." / "But here's where it gets interesting." / "And get this." / "Let that sink in." / "Think about that."
These exist because the writer didn't trust the material. The fix is never a better transition — it is DELETING the transition. Beats connect through momentum: each sentence's claim raises the question the next sentence answers.
SLOP: "Now the setup. It's one command in your terminal." HUMAN: "The whole setup is one command in your terminal."
SLOP: "So think about what this means. The skills standard just won." HUMAN: "That's the part nobody's processed yet. The skills standard just won."
A scene opens MID-THOUGHT, with a concrete subject — the way a person keeps talking after an edit cut, not the way a slide deck starts a new slide. Read the two scenes either side of every cut aloud as one continuous speech: if the join sounds like a section header, rewrite the join.
(Pipeline note: caption alignment matches each scene's first two words in the transcript — so openings must be distinct across scenes, and never start a scene with a digit. That's the only constraint. "So"/"Now" openers were a legacy requirement and are gone.)
3b. SCREEN & SOUND RULES (from what measurably works)
- The FIRST FRAME is the thumbnail. The hook text must be readable from frame one — it is what people see in the shorts feed before pressing play.
- Visuals must match the words being spoken at that moment. Never leave a static or irrelevant screen while talking about something else.
- No music in shorts. It measurably doesn't help this genre; voice + motion carry everything.
- 30-40 seconds is the winning zone. Never let a video feel cut off.
4. THE CLOSE
The CTA line is provided verbatim — never modify it, never add anything after it. The sentence BEFORE the CTA should be the strongest value line in the video (leave them at the peak, then ask).
5. SELF-CHECK before returning (fix, don't flag)
- Read the whole script aloud in your head as ONE person speaking. Any sentence that sounds like a template or a non sequitur: rewrite in plain speech.
- Does every reference resolve? (No "number one" without a list, no "this tool" before naming it.)
- Hard fact in sentence one?
- If there's a loop: exactly one bold claim, and its payoff is the strongest later moment?
- Every number real and spoken as words? Zero invented numbers?
- One screenshot-worthy line per beat?
3c. FORMAT V5 — THE MIX (locked 2026-07-12)
A video that holds one framing loses people. The framing itself is a retention device. Rules:
- NO two consecutive scenes share a framing. Rotate between: FULL FACE (avatar + overlay text), 50/50 SPLIT (animated panel top, face bottom: isen scene with half: true), and FULL ANIMATION (isen/shot full frame).
- The hook is ALWAYS full face with overlay text (frame one = thumbnail).
- Put the most complex visual (org charts, level stacks, faceoffs) in a FULL ANIMATION scene; put reactions, punchlines, and "here's what it means" in FACE or SPLIT so the human carries the emotion.
- A good default arc: FACE hook → FULL animation (the evidence) → SPLIT (face reacts while data stays up) → FACE (the payoff line) → CTA card.
- Never leave any framing on screen longer than ~8 seconds without something inside it changing.
3d. WRITE FOR THE VOICE ENGINE (2026-07-17)
The clone reads punctuation as direction. Long flowing hook sentences come out sounding like AI narration. For energy, especially in the hook:
- Break the hook into short fragments with hard stops.
- A question beat ("Number three? A free CapCut killer.") forces a natural pitch rise and snap that a comma never gives.
- Never open with a 20+ word sentence. Two short sentences beat one long one every time, same words.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

