mr-tbot/bot-code-mods/auto-media-stinger

/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.