agentleFS
raphaelbgrraphaelbgr/ai-video-prompting/skills/viral-short-script/SKILL.md

viral-short-script

Skill0 starsChanged 2 months agoMITView on GitHub

Use when writing or planning the script and structure of a short vertical video (Reel/TikTok/Short, under ~60s) - a visual-first, sound-off scroll-stopper skeleton. Triggers on "viral script", "hook for a reel", "structure of the video", or feeding a video-generation pipeline. Theme-agnostic - it produces the skeleton; the topic is plugged in afterwards.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    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
  1. Distinctive:

    Do this before anything else.

    line 30
  2. Distinctive:

    Every beat is written on two layers at once, because most viewers never turn the sound on:

    line 40
  3. Distinctive:

    Write *line + text + visual* per beat, not a 130-word monologue.

    line 114
  4. Distinctive:

    Write the finished script as a machine-readable object so the render step skips AI planning entirely:

    line 133

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

viral-short-script

A theme-agnostic skeleton for sub-60s vertical video. Structure first; the topic is plugged in afterwards.

Provenance, stated honestly: the story shape (hook → escalate → turn → pay off) was inspired by a structural read of 47 acclaimed films, then stress-tested by a panel of seven frontier models. It is a heuristic, not a proven formula — the film study had no control group and no engagement data. Treat it as a strong starting structure and validate it against your own retention numbers.

The reality it is built around: the feed is hostile, watched muted, and frame 1 is a thumbnail. You do not earn attention — you stop the thumb in about one second with a picture, then hold it with rising tension and a sensory change every couple of seconds.

Use for: anything under ~90s; writing or fixing a hook; feeding a generation pipeline; when a draft trails off, buries the payoff, or reads like a monologue. Not for: long-form (>3 min), carousels, or news briefings with their own planner.


THE GATE — the 3-second silent test

Do this before anything else. Mute the first 3 seconds. From the frame and the on-screen text alone, can a stranger tell who it is for and feel a question or a tension?

No → rewrite frame 1. This single binary gate replaces manual hook-scoring.


The skeleton — 4 beats, two parallel layers

Every beat is written on two layers at once, because most viewers never turn the sound on:

  • SCREEN — what is seen and read: first frame, burned-in text, cut, zoom, sound hit.
  • STORY — what is said, and what happens.

Times are soft defaults for a ~45–60s clip, in absolute seconds — not laws. (Film act percentages do not port to 60 seconds. Ignore them here.)

Beat~TimeJobPass criterion
1. STOP0–2sStop the scrollSurvives the 3s silent test. Frame 1 works as a thumbnail. No "hi guys", no setup shot.
2. STAKES2–8sEarn the next 5sOne line names the payoff or the risk, and for whom. Exactly ONE question is open. A claim carries a proof token.
3. RUN8s → last 8sOne rising arc + interruptsNo plateau. A micro-payoff every ~5–8s. Exactly ONE turn. A sensory change every ~2–3s.
4. LANDlast ~5–8sEmotional punctuation + 1 triggerCloses on emotion (clean snap or open loop), ideally loops back to frame 1. Exactly ONE action trigger.

Beat 1 — STOP

  • SCREEN: frame 1 is motion, a reacting face, a strange object, or a visual contradiction — plus a burned-in text hook. The text may contradict the image for surprise.
  • STORY: show a result-glimpse or a tension, never a preamble. Qualify the viewer implicitly ("if you're the kind of person who…"). Identity and confessional framings beat tired angles like "Stop doing X" or "3 things about X".

Beat 2 — STAKES

One sentence: what is at risk or what they get, and for whom. Install exactly one question. Any claim gets an instant proof token — a number, a receipt, a screen recording, a before/after. Credibility in seconds, not paragraphs.

Beat 3 — RUN (keep the two layers independent)

  • STORY: one tension arc — single problem, escalation with no plateau, then ONE turn that flips the meaning. The turn has no fixed timestamp; place it where it hits hardest. In short-form the turn is the biggest pattern break. Drop a micro-payoff every ~5–8s so someone who leaves early still got something.
  • SCREEN: change one sensory channel every ~2–3s. Menu: punch-in · hard cut · zoom-to-keyword · whip-pan · snap-cut · text pop · sound hit · angle change · colour shift · number-appears · before/after · progress marker (1/3, 2/3, 3/3).

Beat 4 — LAND (close on emotion, not on theme)

  • Made them laugh / gasp / feel satisfied → close clean, with a snap line.
  • Made them curious / angry → leave an open loop.
  • Premium close: loop back to frame 1 so a rewatch feels seamless (lifts average percentage viewed).
  • Exactly ONE action trigger, native to the emotion: share ("send this to the person who…"), save ("save this before you forget"), or comment ("which one did you guess?"). Never a generic "follow for more".

Format modes — pick one; it decides where the payoff sits

ModePayoff placement
story / confessionhold the turn for near the end
myth-busthold the "but actually" for the turn
tutorial / how-tomicro-payoffs throughout, result near the end
demo / transformationfront-load a result-glimpse, full reveal at the end
ranked listprogress markers (1/3…), save-bait close
reaction / commentaryreact first, sharpest take last
news / explainerthe "so what" is the turn
product / advalue first, ask once, late

Also consider platform divergence (Shorts, TikTok and Reels reward different lengths and pacing) and series / Part 2 as a growth lever.


Optimise for the real signals

3-second retention → average percentage viewed / completion → shares (sends per reach) and saves → comments. Iterate hooks against watch-time data, not opinion.

Density budget: short-form is roughly 40% visual beats, pauses and on-screen text. Write line + text + visual per beat, not a 130-word monologue.


ADAPT TO THEME (the deliberately-later step)

  1. Pick the mode from the table.
  2. Name audience / pain / desire, one line each.
  3. Pick the goal (reach vs authority) → which close (loop vs clean).
  4. Fill each beat's SCREEN + STORY, then run THE GATE.

Credibility play for factual topics: source every claim to an official record, include the skeptical counterweight on purpose, and never overclaim. It reads as confidence, and it survives the comments.


Handing the script to a generator

Write the finished script as a machine-readable object so the render step skips AI planning entirely:

{
  "hook": "...",
  "segments": [
    {"text": "...", "keywords": ["...", "..."], "duration_seconds": 4}
  ],
  "cta": "...",
  "total_duration": 48
}

Then feed each segment's visual into the model-specific prompt block — see veo-flow-prompting, wan-3-prompting or seedance-prompting. For the hook shot specifically, generating it is usually worth the money even when the rest is stock footage — the first two seconds are where every cold account looks generic.


Common mistakes

  • A strong first sentence but a weak first frame → dead on arrival in a muted feed.
  • Two competing questions → pick one.
  • A fixed "midpoint reversal at 30s" → that is the retention death-zone. Place the turn by impact.
  • Best line at the start with nothing to climb to → keep a turn for late.
  • A 130-word monologue → short-form is visual + text + pause, not narration.
  • "Resolved, tidy, follow for more" → close on emotion with one native trigger instead.

How to improve this skill

  1. Real feedback loop (highest value): log retention, shares and saves; tag which hooks, closes and modes won; reweight the defaults by measured performance.
  2. Better evidence: re-extract structures with several models and keep only the flags they agree on; add a control corpus (un-acclaimed films + actual viral videos).
  3. A/B: three hooks per topic, ship the variants, keep the winners.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.