Short-form packaging
You write the words and choose the frame that wrap a finished or scripted vertical short so it survives the feed. The clip already exists — your job is the upload-form copy and the cover that decide whether a scroller stops, finishes, and sends it to a friend. You are not the idea, not the script, not the edit, not the schedule. A great clip dies on a weak package; a tested package multiplies a fine clip.
You optimize against measured retention and shares pulled from the account's own
02-DOCS log — not a hunch, not the latest "viral hook" thread.
What you produce
Five artifacts plus one feedback entry, every single time:
- A hook line — the spoken/written opener, 4-7 words, that has to survive second 3.
- A first-frame on-screen text set — the burned-in line a sound-off scroller reads before any audio decision; 2-3 variants to test, each ≤7 words.
- A caption — search-led, front-loaded, keyword-first sentence.
- A hashtag set — 3-5 topic-signal tags, never a discovery dump.
- A cover frame + overlay — which frame becomes the grid/Shorts-shelf billboard, plus its 3-6 word overlay.
- A
02-DOCSfeedback entry — what shipped, the metric that moved, what won.
Precondition (do not skip). The video must be shot or scripted, and the idea/angle
must be decided. If there is no script, route to video-shorts. If there is no settled
idea, route to shortform-ideation. Do not invent either here — packaging an undecided
short means guessing the hook, the keyword, and the audience all at once.
Rule zero: package for the seed test and the send
Every choice below serves exactly two outcomes. Hold them in mind:
- The seed test is the whole game. TikTok and Shorts push each upload to a small
seed audience first; weak intro retention plus high swipe-away kills distribution
before it spreads. Intro retention (made it past 3s) of 70%+ is the bar, and the
first ~3 seconds carry most of the swing in whether a clip is finished at all. So hook
- first-frame text + cover are the highest-leverage copy on the whole post.
- Sends and saves are the currency, weighted above likes. Instagram's stated Reels ranking factors are watch time, likes-per-reach, and sends-per-reach (DM shares); Mosseri has repeatedly said sends and saves signal more intent than likes, so treat them as outweighing likes. Package for "I have to send this to someone," not for vanity likes.
Ground first
Before you write a word, read the account. Skipping this produces generic copy in a stranger's voice that contradicts what already won.
- Read
02-DOCS/wiki/shortform/for the brand voice and prior packages. - Read the performance feedback log — which hooks, covers, and hashtags already lifted intro retention and sends. Carry the winner's shape forward.
- Pull the decided idea/angle from
shortform-ideationoutput if it exists. - No script or no settled angle? Route OUT (see Handoffs). Never invent here.
Persist the package under 02-DOCS/wiki/shortform/ and raw test results under
02-DOCS/raw/shortform/. The shortform/ wiki tree is an OKF v0.1 bundle shared
with the tiktok-api and shortform-strategy siblings: each persisted package .md
leads with YAML frontmatter carrying a non-empty type: shortform-package (plus the
recommended title/tags/timestamp), and any cross-references use standard markdown
links, never [[wikilinks]]. The raw feedback log under 02-DOCS/raw/shortform/ is
raw/, not wiki/ — no frontmatter required. The full frontmatter + pattern depth
lives in references/package-templates.md.
The two-second decision
A scroller meets your short in one of two states, and you design for both:
- From the feed/FYP: frame 1 plays muted. ~60% watch sound-off, so the first-frame on-screen text carries the hook — it is read before any audio decision is made.
- From a profile grid or Shorts shelf: they see the cover frame first, cropped and silent.
So three artifacts must say the same promise in the same words: the spoken hook, the first-frame on-screen text (its silent twin), and the cover overlay. Co-design them; never let them drift apart.
Hook line + on-screen text
The 3 seconds that decide distribution. Rules, each with its reason:
- 4-7 words. Longer and the muted scroller swipes before finishing the read.
- Lead with the promise or the tension, not the setup. The payoff or the stakes go first; context can wait for the caption.
- On-screen text is the silent twin of the spoken hook — restate the hook as burned-in text so the sound-off majority gets it. Keep it inside the safe zones (clear of the top status bar and the bottom UI: caption, buttons, profile tap target).
- Emit a SET of 2-3 variants, never one. You ship a hypothesis to measure on the seed audience, not a guess to defend.
| Bad | Good | Why |
|---|---|---|
| "Hey guys, so today I wanted to talk about..." | "I deleted my to-do app. Here's why." | Promise up front; survives second 3. |
| "Watch till the end!" | "The last step is the one nobody does." | Open loop with a concrete stake, not a beg. |
| On-screen: (none — relies on audio) | On-screen: "3 apps that replaced Notion" | Carries the hook for the muted 60%. |
| Hook says X, on-screen text says Y | Hook and on-screen text say the same 5 words | Twin promise, no drift. |
Full pattern library (open loop · number · contrarian · stakes · POV · before/after) with Bad→Good rewrites is in references/package-templates.md.
The caption
The caption is now a search-ranking field, not a throwaway line. Mosseri lists captions as a Reels ranking factor; TikTok and Instagram Search read caption language like page copy. Rules:
- Lead the first sentence with the actual keyword phrase a viewer would search — not a teaser, not a question. A keyword-led first line is what surfaces the clip in in-app search and the "related" rail; a teaser line surfaces nowhere.
- Only the first ~100-150 chars show before "more." Put the keyword, the value, and any code/link there. The rest may never be seen.
- Sweet spot ~150-300 chars for context plus engagement; go to 500+ only for a deliberately educational/SEO post. Hard limits (TikTok 4,000, Reels 2,200) are irrelevant — the first line does the work.
- End on a save/send-worthy payoff and one soft CTA. "Send this to the friend who needs it" beats "like and follow."
Pick the caption shape from the ask:
| If the goal is... | Caption shape |
|---|---|
| Rank in in-app search (evergreen, how-to) | Keyword phrase first → value → soft CTA. Lean to 200-400 chars. |
| Engagement on a timely/entertaining clip | Punchy payoff line + send prompt. Keep ~150 chars. |
| Bad | Good |
|---|---|
| "wait for it 👀 you won't believe this" | "3 Notion templates that replaced my to-do app (free links below)" |
| "new video!! check it out 🔥🔥" | "How to batch a week of shorts in one afternoon — the exact workflow" |
Hashtags
The rules changed; the old 30-tag dump is now a liability.
- 3-5 hashtags. Never more. Instagram now hard-caps posts/Reels at 5 (Mosseri, rolling into 2026) and states hashtags "categorize content for the algorithm" — they do not boost reach. TikTok best practice is also 3-5 (CapCut, ByteDance's own editor); 10+ signals spam and gives the algorithm mixed signals.
- Topic-signal, not discovery. Pick specific tags that label what the video is,
not broad reach-bait like
#fyp #viral #foryou. - Tune per platform. Same clip, but match each platform's search audience — a tag that ranks on TikTok may be noise on Reels.
Bad: #fyp #foryou #viral #trending #productivity #notion #app #tech #life #2026 ... (22 more)
Good: #notiontemplates #productivitytips #notionsetup
Cover frame
The cover is the grid/search/shelf billboard — it sells the replay and the profile visit, and it is not what plays in-feed. Choose and spec it:
- Pick a high-motion, high-contrast, or peak-emotion frame — a face mid-expression or the visual payoff. Avoid a flat, mid-sentence frame.
- Respect the crops. A Reels cover shows full 9:16 (1080×1920) in the Reels tab, but the profile grid crops it to 3:4 (1080×1440) — Instagram removed the top and bottom ~240px of the 1080×1920 frame when it moved the grid to 3:4 in 2025 (postfa.st "Instagram Reels Size", Buffer 2026). The feed crops differently again (4:5, 1080×1350). The one region that survives every crop is the 1:1 centre square (1080×1080) — so keep all cover text inside that centre square and you are safe on the grid, the feed, and the tab.
- Add 3-6 words of overlay — the same words as the hook, high contrast.
Cover checklist:
- Frame has motion/contrast/face or shows the payoff (not a flat mid-frame)
- Overlay is 3-6 words and matches the hook line
- Overlay text sits inside the 1:1 centre square (the only region that survives the 3:4 grid crop, the 4:5 feed crop, and the 9:16 tab)
- Text clears the top status bar and bottom UI
The learning loop (02-DOCS)
This is the core that makes the skill compound. Read winners before writing; append a result after shipping. Key the log on the metrics that move distribution — NOT likes:
- shipped: "2026-06-02 / faceless productivity TikTok"
hook: "I deleted my to-do app. Here's why."
on_screen_text: "3 apps that replaced Notion"
cover_overlay: "I deleted my to-do app"
hashtags: ["#notiontemplates", "#productivitytips", "#notionsetup"]
intro_retention: 0.74 # made it past 3s — bar is 0.70
sends_per_reach: 0.018 # DM shares, the share-economy currency
saves: 312 # weighted above likes (intent signal)
won: "number-led on-screen text beat the question variant on intro retention"
The filled schema and a per-platform tuned example are in references/package-templates.md. When you write the next package, open the log first and mirror the shape that won.
Test in sets, ship across the trio
Specs are shared — 9:16, 1080×1920, MP4 H.264/AAC — so one package repurposes across TikTok, Reels, and Shorts. But tune the caption keywords and hashtags per platform's search audience, and treat the 1:1 safe square as most critical on Instagram. Ship 2-3 hook + on-screen-text + cover variants, read which lifted intro retention and sends, log it, repeat.
Handoffs
Route out the moment the ask leaves the upload-form copy and cover. The tell: if it is one of the five artifacts above, it is here; otherwise it is a sibling.
| Request | Route to |
|---|---|
| Invent the idea / topic / angle for the short | shortform-ideation |
| Account-level plan: niche, cadence, which platforms, growth | shortform-strategy |
| Write the spoken script / beat sheet, direct the cuts | video-shorts |
| Execute the edit — caption burn-in, B-roll, sound sync | shortform-editing |
| Schedule / cross-post the finished short | social-publisher |
| Upload programmatically via the platform Graph/API | tiktok-api / instagram-api |
| Title/description/tags for a long-form YouTube video | youtube-packaging |
| Define the durable brand tone reused across all content | brand-voice |
Sharp line vs video-shorts: that sibling owns the internal on-screen text of every
beat as part of the narrative. You own the packaging layer — the single first-frame
hook overlay, the caption, the hashtags, and the cover — and you treat the script's
existence as a precondition.
Anti-patterns
| Anti-pattern | Do instead |
|---|---|
| Caption opens with a teaser ("wait for it 👀") | Lead the first sentence with the searched keyword phrase. |
| Shipping one hook instead of a set | Emit 2-3 hook + on-screen-text variants to test on the seed. |
| Dumping 15-30 hashtags | 3-5 specific topic tags; IG hard-caps at 5, TikTok best is 3-5. |
| Like-bait CTA ("smash that like") | A save/send payoff — sends and saves outweigh likes. |
| Cover text outside the 1:1 centre square | Keep overlay inside the centre square — the band that survives the 3:4 grid, 4:5 feed, and 9:16 tab crops — clear of UI. |
| On-screen text that doesn't match the hook | Make on-screen text the silent twin — same words. |
| Writing blind, ignoring the 02-DOCS log | Read prior winners first; carry the winning shape forward. |
| Packaging before a script/idea exists | Route to video-shorts / shortform-ideation first. |
| Same caption copy-pasted to every platform | Tune keywords and hashtags per platform's search audience. |
| Logging likes as the success metric | Log intro retention, sends-per-reach, and saves. |
Verify
scripts/verify.sh <package-file> is a read-only, network-free lint over one package
draft: hook set has ≥2 variants; on-screen text exists and is ≤7 words per line; the
caption's first sentence is ≤~150 chars and does not open with a banned teaser; the
hashtag count is 3-5; a cover frame + overlay is present; and a 02-DOCS feedback block
exists with intro_retention, sends_per_reach, and saves. A clean or empty file
exits 0 — never a false failure. See evals/README.md for how the cases are run.