Short-Form AI Video is a Claude AI skill — live platform research before every script. TikTok, Reels, Shorts. Two variants per run.
You post the same video to TikTok and Instagram Reels. Same footage, same caption, same hashtags. TikTok gets 40,000 views in 48 hours. Reels gets 300 and flatlines. No one can explain why. The common explanation is "the algorithm" — which is technically accurate and practically useless, because the algorithm is just a name for the sum of the platform's completion, engagement, and share signals as applied to your content on that day in that context. The real explanation is that the same script produced different completion rates on each platform because the audiences have been conditioned differently, the feed contexts are different, and the hook format that works on one is already stale on the other.
This is the problem that platform-blind AI scripts accelerate. When you use generic AI to write a short-form video script, you get something that is structurally competent and contextually deaf. It gives you a hook, a middle, and a call to action, all correctly positioned. It has no idea whether the hook format it chose was current last quarter or is current this week, whether the CTA placement it selected matches the completion pattern your specific platform rewards, or whether the caption structure it produced signals quality or spam to the algorithm that routes your content.
Every one of those gaps costs distribution. And distribution is the only metric that matters in short-form.
The Three Signals Short-Form Algorithms Actually Measure
Understanding why platform-specific scripting matters requires understanding what the algorithms are actually scoring. It's not views. Views are an output. The inputs are completion rate, rewatch rate, and share-to-save ratio — and each platform weights these differently, which means a script optimised for one platform's signal mix will perform differently on another's.
Completion rate is the percentage of viewers who watch to the end. It's the most heavily weighted signal on TikTok and the foundation of initial distribution. A script's pacing, its promise fulfilment structure, and the placement of its payoff are all completion-rate variables. If the payoff comes too late for TikTok's audience conditioning, completion drops and distribution dies. The same payoff timing on YouTube Shorts, where the audience has slightly longer patience for build-up due to the browse context, performs better.
Rewatch rate — the percentage of viewers who replay the video — is particularly valued on TikTok for certain content types, especially educational and how-to content where the information density rewards re-viewing. Scripts that include a dense information drop mid-video, or that end with a callback to the opening, can deliberately engineer a rewatch prompt. Generic AI scripts don't know this mechanic exists; they just end at the CTA.
Share-to-save ratio varies by content type and platform. Saves are the dominant quality signal for educational content on Reels — they tell the algorithm the content was worth archiving. A script optimised for saves front-loads the value, delivers a clear summary moment, and places the CTA after the summary rather than replacing it. A script optimised for TikTok shares front-loads relatability or surprise rather than information. These are different scripts with different structures.
Generic AI scripts are optimised for legibility — they're clear, well-structured, and correctly formatted. They're not optimised for any platform's specific completion-signal mechanics, because those mechanics change faster than any model's training cycle.
That gap is exactly what the Short-Form AI Video Prompt skill for Claude was built to close.
What Each Platform Actually Requires From a Script
TikTok, Instagram Reels, and YouTube Shorts are distribution machines with meaningfully different audience conditioning patterns, hook conventions, and CTA logic. Here's what each requires from a script — and where they diverge.
| Platform | Hook window | Pacing pattern | CTA logic | Caption behaviour |
|---|---|---|---|---|
| TikTok | First 1.5–2 seconds. Specific, personal, or counterintuitive statement. "Did you know" and numbered-list hooks are now saturated and depress completion. | Fast-cut, high-energy. Pattern interrupts every 8–12 seconds maintain completion. Build-reveal structure over linear explanation. | Reason-before-ask. "Follow — I'm posting X" outperforms "follow for more tips." Forward incentive, not backward justification. | First line mirrors hook for text-match. Two to four hashtags; over-hashtagging suppresses reach in most niches. |
| Instagram Reels | First 2–3 seconds. Aesthetic consistency and creator identity cues matter alongside verbal hook. Audiences are in a more curated-feed context. | Slightly slower build tolerance than TikTok. Story-first structures perform well. Visual continuity between frames signals production quality. | Profile-reference CTAs outperform generic follows. "Check my bio link" outperforms "link in bio" because it directs an action rather than stating a location. | First line carries more weight than hashtags. Saves are the primary quality signal for educational content — CTA should match: "save this for later" drives the signal. |
| YouTube Shorts | First 2–3 seconds. Shorts audiences arrive from the YouTube ecosystem with more tolerance for topic setup before the payoff — they're in a browsing, not scrolling, context. | Slightly longer build tolerance. Educational density performs well. Chapters or segmented structures create rewatch incentive. | Subscribe CTAs perform better here than on TikTok or Reels because the audience is already in the YouTube subscription mindset. Comment prompts with a specific question drive strong engagement signals. | Titles behave more like YouTube video titles — keyword-relevant and searchable. The caption is secondary to the title for discovery. |
None of this is static. The hook norms in this table reflect current platform conditions — in four to six months, some of these patterns will have shifted as the platforms' audiences continue to condition and the algorithms continue to adjust their signal weighting. That's the fundamental problem with static training data for short-form scripts: the table above would have looked different six months ago, and it will look different six months from now. Live research before every script is the only way to stay current.
Why Format Trends Expire Faster Than Model Training Cycles
Short-form format trends on TikTok and Reels cycle on a two-to-four month cadence for major shifts, with smaller variations weekly. The "controversial statement" hook dominated TikTok through much of 2024 and then saturated to the point where audiences had conditioned a skip reflex to it. The POV format drove massive Reels engagement in 2023 and is now used primarily as a comedy subversion device by creators who know the audience expects to skip it. These aren't minor aesthetic shifts — they're changes in what the algorithm has learned to associate with low-quality templated content.
AI models are trained on data with a cutoff. By the time a model is deployed, the short-form trends it learned from are already ageing. By the time the model has been in use for six months, some of those trends are actively counter-productive. This is not a criticism of how models work — it's a structural mismatch between training timelines and platform evolution rates. The only way to address it is to inject live research into the scripting process before the script is written.
Using last quarter's hook format on this week's TikTok is like sending a cold email with "I hope this finds you well" — technically valid, immediately disqualifying.
The Short-Form AI Video skill addresses this directly. Before producing any script, it fetches what's currently in distribution for your platform and content category — which hook formats are getting high initial completion, which structures are being rewarded in the algorithm's current state, which CTA patterns are producing the engagement signals the platform is weighting. That research informs every structural decision in the script. The result is a script built for the platform as it exists now, not as it existed when a model was last trained.
The Hook Formats That Have Already Expired
One of the most useful things to know about current short-form platform behaviour is which hook patterns have been so saturated that the algorithm now associates them with templated, low-quality content. These are formats that were current and effective — often for good structural reasons — and have since been conditioned out of both the audience's attention and the algorithm's distribution preference.
| Hook format | Why it worked | Why it's expired | What's replaced it |
|---|---|---|---|
| "Did you know that…" | Created curiosity gap immediately. Worked on early-stage audiences not yet trained on the format. | Audiences have seen this opener thousands of times and trained a skip reflex. The algorithm has associated it with mass-produced template content. | Specific personal statements or dollar figures: "I lost $4,000 before I learned this." Stakes are personal, not abstract. |
| "Here are X things you need to know about…" | Clearly signalled value and set expectations. Drove initial completion through a structured promise. | The numbered-list format became synonymous with bulk AI-generated content. Audiences now associate it with low-effort production. | Outcome-first structure: lead with the result, then explain the path. The promise is implicit in the outcome, not a numbered list. |
| "POV: you just discovered…" | Created immersive identity connection. Pulled the viewer into the scenario immediately. | Overuse made it a comedy template — audiences expect irony when they see "POV:" and the earnest use now reads as unaware. | Direct address: "If you're still doing X in 2026, watch this." Confrontational specificity rather than scenario-setting. |
| "Stop scrolling if you…" | Created a pattern interrupt by naming the behaviour (scrolling) and breaking it with an identity qualifier. | The format became so saturated that it no longer pattern-interrupts — audiences read "stop scrolling" as the signal to scroll faster. | Immediate specificity in the first two seconds, no meta-commentary on the medium. The scroll stops when the content is specific enough to feel personally relevant. |
Generic AI scripts still produce hooks from the first two columns of that table regularly — because those formats are well-represented in training data from the period when they were effective. The skill's live research catches this before output reaches you: expired formats are flagged and replaced with what's currently performing in your niche and platform combination.
Common Questions About Short-Form Video Scripting
The creators who grow consistently on short-form aren't the ones with better ideas. Every platform has an inexhaustible supply of good ideas. They're the ones whose scripts are formatted for the platform's current distribution logic — on the day they post, for the audience that's conditioned to that platform's native feel, using hook formats that haven't yet been saturated into algorithmic suppression. That's not a creative advantage. It's an information advantage. And information that has to be current by definition can't come from a model that was trained in the past.
The next piece most people tackle from here is product ad film prompts that brief a director properly. If you're working across the full Video & Pod workflow, the Video & Pod bundle covers everything in one place.
If you're doing this manually, the ai video prompt skill is the faster path. For paid creative specifically, AI UGC ad scripts apply the same platform-aware approach to performance ad content.
Put this to work: the Short-Form AI Video Prompt skill for Claude turns everything above into one guided workflow you run in a normal Claude chat. Not ready to buy? Start with a free Claude skill and see how it works first.