name: auto-media-stinger description: "Use when a project's main stinger — the short launch commercial that plays before anything else — needs writing, cutting, or bringing back in line with what the product now is; when the user mentions the stinger, the launch ad, the trailer, the sizzle reel, the hype cut, the thirty-second spot, or says \"/auto-media-stinger\"; or when the name, price, tagline, platforms or hero screen has moved and the ad still sells the old product." metadata: version: "1.0.0"

/auto-media-stinger

One deliverable: the stinger. The 15–30 second launch spot that plays at the top of the landing page, in the app-store preview slot, in a paid feed placement, and before a demo. One master timeline, three renders, one poster frame, one ship gate.

If the request is "make the videos", that is /auto-media-maker. If it is "make the video — the one that plays first", it is this one.

What This Skill Inherits

/auto-media-maker owns the shared foundation and this skill reads it rather than re-asking: pipeline discovery and .agent/media.json, the avatar and voice, the brand folder and its anti-references, the verbatim never_say list, the catalog, the watch/on_change machinery, per-line content-addressed narration, vN filenames, blessing, and the memory entry format. Its capture hygiene is law here too — scripted drive, no hand-scroll, cursor hidden, never film production writes, never show real customer data or a real card or a placeholder domain.

If none of that exists yet, bootstrap only what a stinger needs and hand the rest back: run the parent's Step 1 pipeline discovery, pick an avatar and voice from the real account listing, get the brand folder, capture the never_say list, and ask the advertisement style set only. Write one catalog entry. Then say in one line: "The rest of the series is unplanned — run /auto-media-maker when you want the catalog." Offer; do not perform.

Never plan a catalog here, never build a second config, memory file or CLI, and never re-ask an interview question the parent already answered.

A Stinger Is A Different Object

Four differences; every decision below falls out of them.

It is the only piece allowed to be striking. A tutorial that shows off has failed — the screen is the subject and the edit should be invisible. A stinger has no such duty. This is the one place in the whole media surface where rhythm, contrast, hard cuts, a real sound-design idea and a point of view are correct. Striking means composed, not busy. But do not import the tutorial's restraint: a stinger that looks like a well-behaved walkthrough has failed at its only job, which is to make someone stop.

It lands one idea; it does not teach. If you cannot say the idea in six words — "your builds stop breaking on Fridays", "one link, every device" — you have a trailer for a product nobody asked to see. Feature lists are the failure mode. Three features in twenty seconds means the viewer remembers zero.

It is judged in the first two seconds — before the audio has finished fading in, and on a scrolling feed before the viewer consciously registers that a video started. It must also survive being a static thumbnail at 200px wide, because in half its placements that is what it is until someone taps.

It is watched on mute. That changes the structure, not just the finishing.

The Beat Map

Written for a 20-second master. Durations scale; the proportions do not.

The final 1.5 seconds must be completely static — no motion, no residual animation, no logo still settling, no motion blur. It is the frame a paused player, a scrubbed timeline and a platform-generated end thumbnail will show, and it is the viewer's only chance to actually read the name and the URL. A logo still easing into place at 19.6s is a logo nobody read.

The mark holds exactly three things: name, one clause, one destination — a domain or a store badge, not both. No social handles, no QR code fighting the URL, no "coming soon" plus a date plus a badge. Anything added there is subtracted from all three.

The first frame must be legible at thumbnail size, must not be a logo, must not be black, and must carry the highest contrast in the piece. A logo-first stinger is the most common failure in product video: it spends the only two seconds you are guaranteed on information the viewer cannot use yet.

Built For Mute

Assume a large share of feed plays are silent. (The widely-quoted 85% figure is old, vendor-sourced and contested — do not cite it. The design rule does not need it: if a meaningful fraction of plays are silent and the spot is incomprehensible silent, it is broken for that fraction.)

The audio track is additive, never load-bearing. Every idea must survive muted. Sound design adds physicality, pace and polish; it may not carry a single fact. No voice-over saying something the screen does not, no punchline delivered in audio, no comic timing that exists only in the mix.

A 20-second launch spot usually should not have a voice-over at all — a talking head is how a stinger degrades into a short bad tutorial, and a talking avatar in one is a different deliverable with a different name. If there is VO, 100% of it is burned in. Not platform auto-captions: those are positioned by the platform, land in the chrome zone, restyle themselves, and are absent entirely when the file is embedded on your own landing page.

Text hierarchy: at most two levels visible at once, one new text object at a time. Primary text is at minimum 1/18 of frame height (~60px in a 1080-tall landscape frame, ~106px in 1080×1920). That is a floor. Secondary text is never below 1/28 of frame height and never below the family's regular weight. Contrast against the worst frame it covers, not a still you chose — hold text over a scrim or a solid unless you have checked every frame underneath.

Dwell time: ~0.3s per word plus 0.4s to settle, minimum 1.2s, longer if the viewer must also look at something else in frame. Read every card aloud at natural speed while it is on screen; if you are rushing to finish before the cut it is too short — and you already know the line, which the viewer does not.

One Master, Framed For The Crop

Master 16:9 at 3840×2160, or 1920×1080 if every source is a 1080 screen capture — never upscale UI. Inside it, define a protected column: the centre 9:16-shaped region, 1080 wide in a 1920-wide frame. Every load-bearing element — the idea line, the mark, the URL, the subject of every shot — lives inside that column for the whole piece. Then 9:16 and 1:1 are crops of the same edit, same cuts, same durations, same audio. If a cutdown needs its own edit, the master was framed wrong; fix it upstream rather than opening a second timeline.

If the product's own UI is portrait, master vertical at 2160×3840 and derive landscape by extraction plus a designed background. Letterboxing a landscape UI into 9:16 with blurred bars is the visual signature of a lazy cutdown and everyone reads it instantly.

Vertical safe areas. Platform chrome eats the bottom third and the right rail. Working numbers at 1080×1920, current as of 2026 and shifting with every app release — a starting point, then verify on a real device:

Design to the intersection: roughly a 900×1400 box centred in 1080×1920 — about 12% to 72% of frame height with ~17% margin each side. The single rule that saves the most re-cuts: never put the logo, the URL or the CTA in the bottom third of a vertical. In vertical the mark goes upper-middle; in 16:9 it goes lower-third. Those are different positions and the crop will not do it for you — build the mark as two positioned variants in the master and switch by output.

Deliverables from the one master: ProRes 422 HQ master; H.264 High Profile 4.0 at 10–12 Mbps VBR with AAC 256 kbps stereo at 44.1 or 48 kHz for store and web; a poster/first-frame still; a designated end-frame still. Pick one frame rate and never mix — 60fps device capture conformed into a 24p cut judders visibly on UI motion, and it is the kind of thing nobody can name but everybody sees.

Loudness

Target for a stinger: −14 LUFS integrated, true peak ≤ −1.0 dBTP. If the same spot must also run as a broadcast commercial, deliver a separate −23 LUFS (EBU R128, −1 dBTP) or −24 LKFS (ATSC A/85, −2 dBTP) version. Do not try to satisfy both from one file; you will satisfy neither. LKFS and LUFS are the same measurement under different standards bodies.

Platform normalization levels: YouTube ≈ −14 LUFS and it never turns anything up; Spotify −14 ("Normal", with −11/−19 user options) and true peak below −1, below −2 if the master is louder than −14; Apple Music Sound Check ≈ −16 with true peak never above −1. TikTok, Instagram, Facebook and X publish nothing — independent measurement lands around −14 to −15, and TikTok's behaviour is disputed and has changed more than once. Do not design around an unpublished number that moves.

Why a hot master sounds worse — the argument to give when someone asks for it louder. Normalization applies negative gain only. Master at −9 LUFS and YouTube applies −5 dB: the viewer hears a transient-crushed mix at exactly the same perceived loudness as everyone else. You paid the whole cost of the loudness war and collected none of the benefit. A −14 master with intact transients sounds bigger at the same normalized level, because the surviving peaks are what the ear reads as impact. Verifiable, not taste: right-click a playing YouTube video → Stats for nerds → content loudness; 0.0 dB means you hit it exactly.

Two short-form traps. Integrated loudness over 15–20 seconds is dominated by its loudest five, and BS.1770 gating behaves oddly on very short programmes containing silence — so measure integrated and short-term (3s) maximum; −14 integrated peaking at −8 short-term over the impact feels shouty and still gets pulled down. And the opening impact is the most common clipping site in the piece — check true peak over the first 500 ms, after encode, not before.

Music Licensing — The Clause That Ends Projects

It has one name: paid advertising (also "paid media", "advertising and promotion").

Catalogues draw the line between a creator tier — content on your own channels, including sponsored and branded content there — and a commercial tier covering advertising, client work and brand campaigns. A stinger sitting organically on your own channel and landing page is often covered by the cheaper tier. The moment a dollar of media spend goes behind it, most creator tiers stop covering it — and the track you already shipped in three aspect ratios becomes an infringement.

As of 2026 and subject to restructuring: Epidemic Sound's Creator plan covers sponsored and branded content on your own channels but not digital advertising for a brand, which needs their Commercial plan; Artlist's unified licence permits monetization and sponsorship from its Personal tier, with higher tiers existing because larger uses carry caps. Verify both against the vendor's current licence text — both restructure roughly annually.

Three axes actually determine coverage, and licence pages usually foreground only one:

1. Media type — organic social / paid social / display / TV / OOH / cinema are separate grants.
2. Term, territory and spend ceiling — many "commercial" licences quietly cap at 12 months, one

country, or a spend threshold above which you must upgrade.

1. Whether the grant survives cancellation — this decides whether cancelling the subscription

un-clears your launch spot.

AI music. Suno requires a paid tier for any commercial use; free-tier output is personal non-commercial listening, which excludes advertising outright. The deeper issue survives paying: it assigns you whatever rights exist and warrants nothing. Combined with the US Copyright Office's position on non-human authorship, a purely generated track may be uncopyrightable — you can use it, and you cannot stop a competitor generating a near-identical one and running it against you. There are also live label suits whose outcome can change terms on tracks already in market. So: screenshot the terms page at generation time, store the prompt, model version and track ID with the project, and prefer generated music for texture and beds rather than for the hook that becomes the brand's sonic identity.

Before shipping, not after — a licenses/ folder holding, for every music cue, SFX, whoosh pack, font, stock shot and any UI sound sampled from the product:

• a licence certificate naming the licensee entity (not a person's account email), the asset ID and the date;
• explicit paid-media coverage if any spend is planned, or a written note that the piece is organic-only plus a trigger to re-clear before any boost;
• for AI assets: the terms screenshot, plan tier at generation time, prompt and asset ID;
• the vendor's Content ID / rights-management clearance, obtained before launch day. A properly licensed track still trips Content ID, and discovering the clearance form exists while the launch video is muted is a bad hour.

If you cannot produce that file for an asset, the asset is not cleared, whatever anyone remembers. Never source from a "no copyright music" upload, a royalty-free compilation channel, or a YouTube Audio Library track without reading that track's current, per-track attribution requirement.

Claims, Disclosures And Store Traps

Superlatives. "Best", "#1", "fastest" are objective claims if a reasonable viewer reads them as factual, and a factual superlative needs a named, dated, methodologically-disclosed source on screen — not in the description, not on the landing page. Puffery is defensible; a measurable superlative is not. The cheapest fix is nearly always deleting it, which also makes the spot better.

Comparative claims naming a competitor. In the US, Lanham Act §43(a)(1)(B) reaches false or misleading statements of fact in commercial advertising — and literally true statements still create liability if presented misleadingly. In the EU, Directive 2006/114/EC permits comparison only where it is not misleading, compares goods meeting the same needs, compares verifiable and material features, and neither denigrates the rival nor takes unfair advantage of their mark. Operationally: a written substantiation file — test method, date, exact versions compared, who ran it — completed before the spot ships, and the on-screen wording matches it phrase for phrase, not approximately.

Pricing on screen goes stale, and stale pricing is a false statement. A price burned into a spot that lives on the landing page in three ratios and inside a store preview needing a review cycle to replace becomes wrong the day pricing changes — and it will. Default: do not show prices. If one must appear, use "from $X" with an on-screen "as of <month year>", and record it so a pricing change is a hard re-cut trigger. Same for free-trial lengths, plan names, seat counts, "over N users" counters, funding mentions and supported OS versions.

Testimonials. FTC Endorsement Guides, 16 CFR Part 255, revised 29 June 2023: "clear and conspicuous" means difficult to miss and easily understandable, and a disclosure for a visual representation must itself appear visually. In a muted vertical that means burned in, readable, held long enough to read — never a caption line, never the description, never a six-frame flash. Disclose material connections. Actors portraying users must be identified as such. Atypical results require the typical result stated, not "results may vary" mouse type.

Meta ad policy evaluates ad copy, creative, landing page and comments together — a policy-clean spot pointing at a non-compliant page still gets rejected.

App Store. Guideline 2.3 requires metadata to reflect the app's core experience and stay current. App Previews must use footage captured on device — your cinematic stinger is not an App Store preview. They are two deliverables; a spot full of designed B-roll and composited UI will be rejected, so plan the device-captured 15–30s variant from the same script or accept that the store slot gets its own cut. 2.3.9: you must own rights to everything shown, and account data must be fictional. 2.3.10: no names, icons or imagery of other mobile platforms — the "available on iOS and Android" end card that is correct on your website gets the preview rejected, so build a store-variant end frame. Specs: 15–30s, up to 3 per localization, ≤500 MB, H.264 High 4.0 at 10–12 Mbps or ProRes 422 HQ, AAC 256 kbps stereo, default poster frame at 5s.

Google Play wants a 30–120 second full YouTube URL, public or unlisted, not age-restricted, monetization off so ads do not play before your own promo, with a feature graphic present for it to display at all. A 20-second stinger is disqualified by length — decide this at the beat-map stage, not at upload.

Accessibility, now also a legal surface (European Accessibility Act obligations from June 2025): captions burned in, text contrast ≥4.5:1 against the worst frame it covers, and no more than three flashes per second (WCAG 2.3.1). A bold, VJ-inflected stinger will fail the flash rule by accident. Check it deliberately.

Two absolutes: every identifiable person on screen has a signed release, and every screen shows fabricated data — no real customer, no real card, no real domain, not blurred, not for four frames.

What check Cannot See

A stinger does not go stale when a file changes. Its watch list is nearly file-blind, and that is its defining property: it goes stale on positioning — the product name, the price, the tagline, the one-line pitch, the hero screen, the platforms claimed, store availability. Point watch at marketing copy, pricing config, store listings and the landing page, not at implementation.

So the parent's `check` will usually report the stinger clean while it is materially wrong. Its real staleness review is a human question asked on every run: is this still what the product is? The trigger is a marketing event, not a diff. Register it with exactly those triggers and stop.

The Ship Gate

Thirteen checks, run on the delivered files, not the timeline. Any failure is a re-cut.

1. Mute test. Silent, on a phone, at arm's length. Can a stranger say what the product is and who it is for?
2. Two-second test. Freeze at 0:02 — is the idea legible? Export the first frame at 200px wide; does it still read? Is it a logo or black?
3. Last-frame test. Pause on the final frame: name, one line, one destination, completely static.
4. Crop test. Overlay the 9:16 and 1:1 masks across every frame. Nothing load-bearing outside the protected column.
5. Chrome test. Post to a private slot and screenshot on a real device with the platform's real UI on top. Templates lie; rails move between app versions.
6. Loudness. Delivered file: −14 ±0.5 LUFS integrated, short-term max within ~4 LU of integrated, true peak ≤ −1.0 dBTP including the first 500 ms. Unlisted YouTube upload should read near 0.0 dB content loudness.
7. Dwell test. Every card read aloud at natural speed while on screen.
8. Claim file. Every on-screen factual statement has a line in the substantiation doc with source and date, and the wording matches.
9. Licence file. A certificate for every asset, covering paid media if any spend is planned.
10. `never_say` pass — applied to spoken copy, on-screen text, and anything legible inside captured UI. Screen copy is copy.
11. Stale-data pass — prices, plan names, counters, dates, OS versions, any UI redesigned since capture.
12. Store-variant pass — device-captured and badge-free for App Store; ≥30s and monetization-off for Play.
13. The one-idea test. Someone who did not make it watches once, unpaused, then says the idea back. Two things or nothing means re-cut. This is the only check that catches a technically flawless spot that says nothing.

And the inherited law: never bless a build you did not watch, end to end, with sound. Then watch it again with sound off, because that is how most of its real plays happen.

What This Skill Must Not Do

• Plan a catalog, or accept "can we get a few more like this" — that is a campaign, and it is /auto-media-maker's scope. Saying so is the correct answer.
• Treat cutdowns as new pieces. They are renders of the same master.
• Re-ask the pipeline, avatar, voice, brand or never_say interview.
• Stand up a second config, memory file, task list or CLI.
• Put a talking avatar in a 20-second launch spot.

Rationalization Table

Red Flags — Stop

• Shipping with any asset that has no licence certificate naming the licensee — or with a creator-tier track behind paid media
• An on-screen superlative, comparison, price or count with no dated line in the substantiation file
• Delivering the stinger as an App Store preview without a device-captured, badge-free variant
• A final frame that is still moving, or a first frame that is a logo or black
• Blessing without watching it muted, on a phone, with the platform's real chrome on top