| name | tiktok-script |
| description | Write a TikTok video as a production-ready script — hook, on-screen beat table, voiceover, caption, hashtags — in a voice native to TikTok rather than repurposed from a blog. Use when the user says 'TikTok video,' 'TikTok script,' 'write a TikTok,' 'TikTok hook,' 'myth-bust video,' 'build in public video,' 'screen recording video,' or wants TikTok as a channel for a product. Prefer this over the upstream social skill whenever the deliverable is one actual TikTok script; social owns multi-platform social strategy, LinkedIn/X/Facebook and scheduling. Writing only — it does not upload; to publish see tiktok-content-posting. For a paid TikTok or Meta ad hook see ad-creative. Covers the contrast hook, the beat table, VO word budget, caption craft, burned-in-text collisions, and why engagement bait is a net loss. For Reels see instagram-post, for Shorts see youtube-shorts — the three are NOT the same script. |
| metadata | {"version":"1.1.0"} |
TikTok
TikTok is where a product gets discovered by people who were not looking for it. That single fact sets everything else: no prior context, no goodwill, no patience, and no search intent to inherit. The video has to earn its own audience from a cold scroll, every time.
The mistake that kills most brand TikTok is treating it as a distribution surface for a message written elsewhere. A TikTok is authored for TikTok, and then — if it works — the idea travels to the other channels, adapted. Not the file, and not the caption.
Read first
If .agents/brand-context.md exists, read it — the claim ledger, prohibited-language list and
channel CTA policy. If it does not, do not stop: derive what you need from the codebase as you
draft, and run claim-check on the finished draft, which builds the ledger properly.
What a TikTok script is
A deliverable, not a paragraph. Produce all six parts:
**id** `w{n}-tiktok-{topic}-{format}-{variant}` · **segment** · **gate** · **runtime**
**Asset:** what footage exists, or precisely what must be shot
**Job:** the one thing this video is for
**On-screen beats:** table of in / out / text
**VO:** the spoken track, at the word budget below
**Caption:** the written post
**Hashtags:** 1–2 descriptive (platform hard cap: 5)
Include the gate: SHIP, NEEDS-ASSET, or blocked-on-review. A script whose footage does not exist is a production brief, and calling it ready wastes a slot in the calendar.
The hook
You are interrupting a scroll, and the interruption has to happen immediately — before the viewer has decided anything, including whether to decide.
One correction worth making, because the whole industry repeats the wrong version: TikTok has never published a numeric hook threshold. The "first 3 seconds" framing is a useful principle, but the signal TikTok actually documents is completion — watch completion and time spent are the most heavily weighted signals, and TikTok states explicitly that neither follower count nor prior video performance are direct ranking factors. Write for the finish, not for a three-second gate. (And ignore the "users decide in 1.7 seconds, per Meta's internal data" line entirely — it circulates with no traceable source.)
The hook shapes that work for a product, roughly in order of reliability:
| Shape | Form | Why it holds |
|---|
| Negation | "We didn't build a translator." | Contradicts the category the viewer just assumed |
| Self-indictment | "Our own homepage is lying about our product." | Nobody scrolls past a confession |
| Reframe | "A wag means aroused. Not happy." | Corrects something the viewer believes right now |
| Visible test | "Run the same photo twice." | Promises a result within the video's runtime |
| Concrete mechanism | "Which 4 frames of your video the model actually sees." | Specificity reads as insider access |
All five are statements, not questions. A question invites the viewer to answer it internally and keep scrolling; a statement they disagree with holds them until it resolves.
Three things fire at once in the first beat — visual, verbal, on-screen text — and they should say the same thing rather than three different things.
Build-in-public is a genuine TikTok format, not a fallback. A screen recording of your own editor, your own dashboard, your own copy being deleted, is the cheapest footage in any plan and the lowest claim-risk footage available: the claim is literally on screen, so it needs no external evidence. No animal, no talent, no location, recordable in an afternoon.
The beat table
Write the video as timed beats. This is the artifact an editor works from and the thing that stops a script becoming a wall of prose.
| in | out | text |
|---|---|---|
| 0.0 | 2.0 | WE DIDN'T BUILD / A PET TRANSLATOR. |
| 2.0 | 5.0 | PHOTO OR / SHORT VIDEO. |
| 5.0 | 7.5 | MOOD SCORE + / VISIBLE CLUES. |
| 7.5 | 10.0 | FOR CURIOSITY. / NOT A DIAGNOSIS. |
| — | — | persistent footer: BRAND · AI-GENERATED VISUALS |
Rules that come from the format rather than from taste:
- Two lines maximum on screen, three to five words per line. The
/ marks the line break so the editor does not guess.
- A beat is 2–4 seconds. Faster than that and it cannot be read on a phone; slower and it stalls.
- Persistent footers carry the brand mark and any mandatory disclosure (AI-generated visuals, review status). They are not beats and should not compete with them.
- If the master already has burned-in copy, do not add native text stickers. Note the safe rectangle in the script (e.g.
x=108–864, y=240–1440 on a 1080×1920 master) so nobody places a sticker into a collision.
Voiceover budget
Roughly three words per second of runtime. A 10-second video is about 30 words; a 30-second video about 90. This is a hard budget, not a target — over it, the delivery has to rush and the retention drops with it.
Speak slightly faster than conversation, vary the tone, and pause before the beat you most want remembered. If the video is faceless, an original recording still beats a synthetic voice for a brand that is trying to be believed.
State explicitly whether the VO is mixed into the delivered file or is a future recording script. A deck that blurs the two ships a silent video with a caption promising narration.
The caption
TikTok captions are read after the video, by people who already engaged, which makes them the wrong place for a hook and the right place for the argument.
- First ~150 characters are visible before "more". Put the claim there, not a wind-up.
- Then a short essay in the brand's voice: the reasoning, the constraint, the thing that did not fit in 30 seconds. Two to five short paragraphs.
- Every number traceable. If the caption says the cap is three, that cap is in the code and the script says where.
- Close with the boundary statement and the CTA.
CTA: "Link in bio." Caption URLs are not the intended click path on TikTok and treating them as one produces attribution that does not reconcile. One bio link, held stable for the run, so post-level analytics remain interpretable.
Hashtags
Hard cap: five. TikTok imposed a five-hashtag-per-post limit in early August 2025, delivered as an in-app "Maximum 5 hashtags" popup rather than a formal announcement. Past five, only the first five count for categorisation.
Use one or two, and make them specific. Metricool's 2026 TikTok study (2.3M posts, 92K accounts) found hashtag-driven views up 114% year over year, and posts carrying at least one hashtag getting roughly 5% more views and 9–10% more interactions. So the "hashtags are dead" line is wrong — but the win comes from descriptive tags that tell TikTok what the video is about, not from volume. Never #fyp or #viral; they categorise you as nothing.
Leave auto-captions on. TikTok uses them to understand the video, and they add accessibility. Burned-in captions are for the viewer; auto-captions are for the system. Do both.
Length: measured, not asserted
The advice to "go longer in 2026" is folklore. It traces to blog explainers with no primary TikTok source behind them, and the only platform statement in that direction is a promotional post from 2021.
What the data says: Metricool's 2026 study measures average TikTok video duration at ~46 seconds, unchanged year over year. Over the same window views fell 31%, reach 29%, and total time watched 87% — while full watch rate fell only ~10% and duration stayed flat. That combination means the decline is a distribution problem driven by saturation (video volume up 72%, image posts up 140%), not a retention problem and not an algorithmic punishment for length.
Practical reading: length is not a lever in either direction. Attention density is. Pick the length the content actually needs, and measure your own completion.
On search, honestly
TikTok search is real and oversold in equal measure. TikTok's own advertising materials cite 57% of users using search and 23% searching within 30 seconds of opening the app — but Metricool measures search at roughly 4% of impressions for an average video, rising past 6% for accounts over a million followers. For You still drives about 73% of views.
So: write for the feed, and let search be a bonus. The "TikTok is replacing Google" narrative is not supported — in a Feb 2026 consumer survey, TikTok ranked last among platforms people find helpful for searching (Google 85%, Reddit 29%, ChatGPT 26%, YouTube 24%, TikTok 16%), and the share of Gen Z saying they lean on TikTok over Google halved from 8% to 4% since 2024.
Educational content is still worth making — it compounds through saves and rewatches, both high-weight signals — but make it because those signals are real, not because of the engagement-rate percentages floating around, which contradict each other wildly depending on which vendor you read.
What TikTok itself says it wants in 2026
From TikTok's own Next 2026 trend report — three trend signals under the year's theme, "Irreplaceable Instinct". The labels are TikTok's; the readings are mine:
- "Reali-Tea" — unfiltered stories, real process and real people over curated perfection. Read: build-in-public and screen recordings are not a budget compromise.
- "Curiosity Detours" — users arrive with intent and leave somewhere adjacent. Read: show up in contexts next to your category, not only inside it.
- "Emotional ROI" — impulse loses to intention, and purchases get pre-validated in comments and by creators. Read: explain the payoff rather than pitch the product.
Platform facts in this file were verified 2026-08-26. Caps and norms move; re-check before betting a quarter on them.
What not to do
- Engagement bait — but know which half of that rule is actually true. "Tag a friend" is bait anywhere: it asks for an action that manufactures distribution rather than expressing interest. But "asking a question" is not the same thing, and the pack's own source says so — Metricool's 2026 study found posts containing a question get 26% more comments. The honest rule is about what you ask: a question that invites someone to say something real is good craft; "which one are you", "comment below", and anything whose answer is a single word are bait, because the comment section they build is worthless to you and to the next reader. And if a comment asks something you cannot answer responsibly, the reply is what is observable plus the fact that the video cannot settle it — nothing more.
- A repurposed horizontal video. Vertical or nothing.
- Generated media presented as evidence. A labelled diagram is fine. A generated photograph passed off as a real example is a fabrication, and the caption does not fix it.
- A hook that takes four seconds to arrive. If the first beat is context, cut it and start at beat two.
- Certainty you have not earned. TikTok rewards confidence, which is exactly why an unearned claim travels furthest before it is checked.
- Trending audio that fights the message. Original or platform-library audio is the safe default for a brand account, and mandatory if the video will be reused as an ad.
Iterating
Test one variable at a time, hook first — same body, different opening, both posted. Hook accounts for most of the variance in reach and almost none of the production cost.
Read hold rate and completion before reach. Reach is downstream of both and lags by days. High views with low follows means the content and the account are mismatched; high saves with low shares means it is useful but not social.
When a hook wins, the correct next move is not to repost it — it is to expand it into a longer reviewed piece and a canonical written answer, so the attention it captured has somewhere durable to land.
Related skills
- instagram-post — Reels look identical and are written differently; do not cross-post the file and call it done.
- youtube-shorts — same aspect ratio, opposite discovery model (search, not scroll).
- claim-check — run on the finished script, including on-screen text and VO.
- tiktok-content-posting — the publishing step, once the script is approved and the asset exists.
- video / ad-creative (upstream) — production tooling and paid-format taxonomy.
