kiss-cam

kiss-cam

A two-call pika pipeline: spectator-POV Jumbotron still (gpt-image-2) → 15-second in-arena kiss cam clip with PA-announcer commentary and crowd reaction (kling-v3-omni, first-frame-locked to the still). The trend look is calibrated — pass both reference images straight through. There is no textual substitution into the prompts; the subjects are anchored only through images. Don't reach for ${subject_a} / ${subject_b} placeholders. Step 1 and Step 2 prompts are verbatim, not scaffolds.

Prerequisites

pika MCP available in the host. Tool name prefix varies by mount point — use whatever the host exposes. Tools needed: identity balance, asset upload, image generation, reference-video generation, and async status follow-up when a generation does not complete inline.

Cost transparency gate

Before any paid MCP call, call identity_balance({verbose: true}) once. Surface the current balance, recent burn rate, and remaining runway, then gate the run with this exact message:

Estimated cost: about 1,500-3,500 credits (~$15-$35) for the GPT-image-2 Jumbotron still, one or two Kling v3-omni pro 15s renders (includes one Step 2 corrective retry with a changed payload), and post-flight analyze_media QA. This exceeds $5, so Reply proceed to continue or cancel to stop.

Do not call any paid MCP tool until the user replies proceed. If the user replies cancel, stop without generating. This is the only yes/no gate; after proceed, the pipeline runs end-to-end.

Pre-generation wall-clock guard

Start a timer at skill start once both subject reference URLs are resolved and the cost gate has passed. The first paid generation call is Step 1 generate_image_edit, and it must be invoked within 5 minutes of skill start. If you have not invoked generate_image_edit within 5 minutes of skill start, stop before any paid generation call and report failed_pre_generation_timeout with what you have so far: subject URL status, upload status, cost-gate status, Step 1 prompt readiness, and the exact blocker. Do not keep refining the Jumbotron prompt, camera language, or style-lock wording.

Print a single-line progress checkpoint after each prep stage and right before the paid generation call:

• Stage 1/3 done — subject references resolved and uploaded.
• Stage 2/3 done — cost gate passed, locking the Jumbotron still prompt.
• Stage 3/3 done — still prompt locked, calling GPT-image-2 now.

Prompt and still-check wording iteration is maximum 2 passes before Step 1. After the max 2 passes, ship what you have to generate_image_edit; do not continue polishing scoreboard details, arena atmosphere, or kiss-cam graphic language.

Long-running task_status polling

When any long-running generation call returns a task_id with or without an initial status, including {task_id}, {task_id, status: "queued"}, or an initial queued, running, or processing status, record the task id and start time immediately.

• Call task_status({task_id}) in a tight loop until terminal (completed | failed | cancelled). No manual sleep and no Bash polling; the worker holds each status call open.
• Emit ONE visible progress line every 60s while status is queued, running, or processing: Seedance i2v queued for {N}m {S}s... still processing. Replace the provider/stage label when polling GPT-image-2 or Kling tasks.
• On completed, unwrap the returned result URL and continue.
• On failed or cancelled, surface failure to the user with task_id, status, and the last status message.
• After 15 min total from the original submit, call task_cancel({task_id}) if the task is still non-terminal, then surface failure to the user. If cancel reports the task is already terminal, call status once more and report that terminal result.
• Do not submit a duplicate request while the original task is still queued, running, or processing.

Stage 0 — Intake

If invoked with empty args and no usable prior context, print this menu and stop:

Which two subjects should be on the Kiss Cam? Required:

• Subject A reference photo — local path or HTTPS URL
• Subject B reference photo — local path or HTTPS URL

If one photo is already present, ask only for the missing photo. Running before both have arrived leaves Step 1 with a missing images entry and produces an inconsistent still.

• Subject A reference photo (required) — local path or https URL. Save as state.subject_a_url.
• Subject B reference photo (required) — local path or https URL. Save as state.subject_b_url.

For each: if already an https://… URL, use it as-is. If local path → upload_asset → PUT bytes to presigned_url → use public_url. On Claude Desktop, pasted inline images don't reach MCP — ask once for a URL or a .zip attachment instead (this is the one allowed clarifier; once both URLs are in, only the cost gate remains).

Either subject can be in any visual style — photoreal human, 3D rendered character, designer toy, illustrated avatar, sculpted figurine, etc. The recipe preserves whatever style the reference uses; do not redraw in a different style. No names are used anywhere — the Kiss Cam graphic does not have a chyron with names. Just two subjects caught on the Jumbotron.

Run the Cost transparency gate, then confirm back in one line ("Generating a Madison Square Garden Kiss Cam moment for these two…") and start. No further yes/no gates after the cost gate — the pipeline runs end-to-end.

Step 1 — Spectator-POV Jumbotron still (generate_image_edit, gpt-image-2)

The kiss cam graphic + scoreboard + retro frame get baked into the still at frame 0 — load-bearing, so Kling treats the entire decorative UI as pixel-locked burned-in UI in Step 2 instead of animating it mid-clip.

Why gpt-image-2 (and no fallback): sharper LED panel detail (scoreboard numerals, kiss cam typography, retro decorative edges) and stronger reference-likeness lock than alternative providers; the LED-sharpness + likeness combo is what sells the trend. On a moderation_blocked response, re-roll the same call instead of swapping providers — alternatives produced softer likeness and softer LED detail in earlier trials. We call gpt-image-2 at 1K 16:9 (1792×1024); higher-resolution variants don't help here since Kling pro outputs 1080p downstream.

Retry budget: Step 1 gpt-image-2 still generation gets at most 3 total attempts, including moderation hits and self-check re-rolls. moderation_blocked counts against this Step 1 cap. Track state.step1_attempt_count before every paid still call.

Why a Jumbotron-POV phone shot (and not a TV broadcast overlay): the first iteration produced a TV broadcast cutaway with a pink-heart kiss cam graphic on the feed — user feedback was "the kiss cam graphics is ugly, look how real kiss cam moments look in real videos." Real viral kiss-cam clips online are virtually all spectator phone shots OF the Jumbotron (Obama-era USA Basketball kiss cam, Sarah Hyland / Wells Adams kiss cam, etc.). The Jumbotron-shot framing hits the aesthetic users actually associate with "real kiss cam" — retro red border + sparkly hearts + cursive Kiss Cam script + adjacent LED scoreboard panels + arena darkness + fans filming with phones.

prompt (verbatim, sent as-is — no template substitutions):

A high-resolution spectator-POV phone screenshot from inside packed Madison Square Garden during a Knicks vs Bulls NBA game, filmed by a fan in the lower bowl looking up at the giant suspended Jumbotron displaying the in-arena "Kiss Cam" segment. Real fan-filmed phone footage aesthetic — TikTok / Instagram online kiss-cam clip style.

Foreground (bottom 20% of frame): silhouettes of the backs/heads of a few fans in the row in front, slightly out-of-focus, dark; a couple of phones held up filming the Jumbotron, bright phone screens visible.

Mid-ground / upper 80%: the Jumbotron dominates the upper-center, bright LED panels against the dark arena ceiling and cabling above. Slight phone-camera tilt, slightly off-center natural handheld framing. Dim arena around, hint of upper-deck silhouettes.

The Jumbotron's central LED panel displays the kiss cam segment in iconic retro arena kiss cam style:
- Thick deep crimson red glowing decorative border frame.
- Bright red sparkly cartoon heart icons glowing in the corners and along edges.
- Subtle white star pattern texture in the red border.
- At the bottom center, large flowing white cursive script "Kiss Cam" with red glow and drop shadow.

Inside the kiss cam frame on the Jumbotron, two subjects framed two-shot chest-up:
- LEFT (Subject A) is the character from the FIRST reference image — preserve their exact likeness AND exact visual style (medium, level of stylization, color palette, rendering attributes). Do not redraw in a different style. Subject A is smiling shyly with one hand near the face, looking caught-off-guard and giggling.
- RIGHT (Subject B) is the character from the SECOND reference image — preserve their exact likeness AND exact visual style. Do not redraw in a different style. Subject B is seated upright at human scale next to Subject A, head tilted slightly toward Subject A.
- Around them inside the frame: packed Knicks fans in blue-and-orange jerseys, laughing, pointing, smiling at the kiss cam.

Adjacent LED panels on the Jumbotron show: "KNICKS 57 — BULLS 61", Q4, clock "4:32", with team logos. Small "MADISON SQUARE GARDEN" / "MSG" / AT&T branding.

Phone-camera aesthetic: slight motion blur, mild handheld imperfection, slightly noisy in dark areas, bright LED slightly blown out, 16:9 aspect ratio, real spectator POV. Looks like a still pulled from a fan-filmed Instagram / TikTok kiss cam clip.

Call params:

• provider: gpt-image-2 (load-bearing — sharpest LED detail and strongest reference-likeness lock; no fallback provider, re-roll the same call on moderation hits only while Step 1 budget remains)
• images: [state.subject_a_url, state.subject_b_url] (order matters — Subject A must be index 0, Subject B index 1; the prompt's "FIRST / SECOND reference image" refers to array index)
• aspect_ratio: 16:9
• quality: medium (default for speed; high is now exposed but ~2 min/call — use only when fidelity matters)
• output_format: png

Do not pass textual feature descriptions for either subject. The prompt above already refers to each subject only as "the character from the FIRST/SECOND reference image" — that's intentional. Describing hair / face / clothing / accessories in text fights the reference image and causes drift: verbal features override the visual reference, and the model homogenizes the subjects toward the description.

The reference images carry all identity + style information; the prompt only adds the style-preservation lock. This works for any reference style — photoreal, 3D rendered, sculpted, illustrated, etc. — without naming the style.

Save the returned URL as state.kisscam_still_url.

Agent-side self-check before Step 2: visually inspect — if the "Kiss Cam" text is misspelled, the scoreboard looks wrong, or either subject's likeness drifted, re-roll Step 1 within the Step 1 cap. Everything downstream pixel-locks to this image. This is the agent's own check — do not ask the user.

On failure (moderation_blocked from gpt-image-2 — often female/female subject pairings + "kiss cam" wording): re-roll the same call only while Step 1 budget remains. Do NOT swap providers — alternatives produce softer likeness and softer LED detail.

Step 2 — In-arena kiss cam clip (generate_reference_video, kling-v3-omni)

Kling-omni image_types: ["first_frame"] locks the still as literal frame 0, so the Jumbotron, scoreboard, kiss cam frame, hearts, cursive "Kiss Cam" text, and foreground fan silhouettes all stay pixel-locked across all 15s. Only the content inside the kiss cam panel animates.

Call params:

• provider: kling
• kling_model: kling-v3-omni
• duration: 15
• aspect_ratio: 16:9
• quality_mode: pro (1080p)
• reference_images: [state.kisscam_still_url] (use the latest value — Step 1 re-rolls overwrite it)
• image_types: ["first_frame"]
• sound: true
• prompt_adherence: strict
• negative_prompt (verbatim):

scene cuts, scoreboard changing, Kiss Cam text changing, graphics morphing, character turning real, character becoming 2D, identity drift, gimbal stabilized, exaggerated acting, theatrical expressions, over-acting, mugging at camera, cartoon reactions, subjects lip-syncing PA announcer dialogue, characters mouthing the announcer lines, subjects' mouths moving with off-screen voice, subjects talking to camera, on-screen lip-sync

prompt_adherence: strict paired with the full negative_prompt anchor list are load-bearing — without both, Kling animates the scoreboard or "Kiss Cam" text mid-clip and subjects regress toward over-acting / lip-syncing the off-screen PA announcer.

prompt (verbatim, ~2400 chars — keep it pre-trimmed because Kling caps prompts at 2500 chars):

First frame: spectator-POV phone shot at MSG of the Jumbotron Kiss Cam segment. Fan-filmed TikTok / Instagram kiss-cam clip style. Handheld phone, NOT pro broadcast.

Camera: continuous handheld shot. Subtle micro-wobble. Slow smooth push-in zoom across 15s toward the kiss cam panel.

Locked: Jumbotron, scoreboard (KNICKS 57 / Q4 4:32 / BULLS 61), red kiss cam border, sparkly hearts, cursive "Kiss Cam" script, AT&T branding, dark arena, foreground fan silhouettes — all consistent.

Inside the kiss cam frame:
Both subjects are the two characters already shown on the kiss cam panel of the first-frame reference still — preserve their exact likeness AND exact visual style across all 15 seconds. Do not redraw either subject in a different style. They animate with subtle, restrained, true-to-life motion within their original style — small natural blinks, gentle breathing, slight head turns, brief micro-smiles. Neither stiff and frozen, nor over-acting. The level of expression is what a real person shows when caught unexpectedly on a stadium camera. Keep all movement subtle, believable, human.
Surrounding Knicks fans react naturally in the background.

Timeline (all reactions stay subtle and human-scale):
0-3s: Subject A notices the camera, small embarrassed half-smile, glances at Subject B. Subject B notices a beat later, soft natural realization, slight smile beginning. Fans behind react gently.
3-7s: A and B exchange a brief warm look, share a quiet small laugh. Faces relax into natural smiles.
7-11s: A and B lean in and share a gentle kiss on the lips — brief, sweet, natural, not staged. Crowd cheers softly.
11-15s: They pull apart with light natural smiles. B leans head gently on A's shoulder. Both share a small quiet laugh.

Audio (non-diegetic / OFF-SCREEN only — Subject A and Subject B stay SILENT throughout, mouths closed, do not lip-sync the dialogue below):
- Packed-arena ambient throughout.
- OFF-SCREEN gender-neutral PA announcer (NOT from either visible subject — an unseen arena voice), amplified, slightly echoey, playful.
  0-3s: "Oh, kiss cam at the Garden — look at this!"
  3-7s: "Hahaha — let's see if they go for it!"
  7-11s: crowd erupts with "AWWWWS," cheers, claps swelling.
  11-15s: PA laughing: "There it is! Big love at the Garden tonight!"
- Phone mic picks up nearby fans louder than PA echo.

Aesthetic: real spectator phone-shot, noisy darks, slightly blown LED, subtle motion blur. 16:9, 1080p.

Why "subtle, restrained, true-to-life motion within the reference style" (and not "subtle motion only", a specific style label, or an expressive micro-expression list):

• First iteration constrained stylized subjects with "subtle head/eye shifts only — remains a vinyl toy throughout" — looked frozen and pasted-in.
• Second iteration named a specific style ("Pixar-quality 3D character"), which forced subjects toward that look even when the reference was a different style.
• Third iteration said "animate naturally and expressively" with a loaded list of micro-expressions ("eyes widening, mouths opening for shock, cheeks lifting when laughing, hands coming up to the face") — subjects then over-acted, mugged at the camera, became theatrical.

The working framing keeps the style-preservation lock but specifies subtle, restrained, true-to-life motion at the level of someone actually caught on a stadium camera — paired with exaggerated acting / theatrical expressions / over-acting / mugging at camera / cartoon reactions in the negative_prompt to suppress regression toward the third-iteration failure mode.

Save the returned video URL as state.kisscam_video_url. If generation completes asynchronously, follow the MCP tool's returned status handle. Client-layer timeouts can create an orphaned upstream task when no task_id reaches the agent; do not submit a duplicate. Surface the timeout and the orphaned upstream task risk to the user, then wait for a recoverable task handle or explicit operator confirmation before any rerun.

Step 2 Kling video generation gets at most 2 total attempts (initial render + one corrective retry for text/scoreboard animation, identity drift, kiss timing, PA timing, or lip-sync artifacts). kling-v3-omni has no seed, and identical Kling payloads can resolve to the same job/asset. Do not submit an identical Kling payload just to seek variation. Before the corrective retry, materially change the payload by using an updated state.kisscam_still_url, restoring missing strict params / negative_prompt entries, or changing the PA audio wording / timing. Track state.step2_attempt_count.

On failure: use the Step 2 corrective retry only after changing the Kling payload — don't switch video engines. Seedance's two-stage likeness gate (same as the baseball-trend sibling) makes it unusable here. If the still itself is the issue (use the Step 1 self-check criteria — text spelling, scoreboard, likeness — to decide), re-run Step 1 only if Step 1 budget remains and use the new still as the changed first frame. After either cap is exhausted, stop and ask for safer references or permission to deliver the best attempt; include the best still/video URL and the failing check.

Step 3 — Deliver

Return both Pika CDN URLs: the still image URL and the final video URL. If the host client requires local media markers, create the local preview outside this skill after confirming both CDN URLs are reachable.

One-line summary: "Kiss Cam moment at MSG — 15s, 16:9, 1080p, Kling v3-omni, native PA-announcer commentary and crowd reaction."

Post-flight quality gate

Before declaring success, call analyze_media on state.kisscam_video_url and ask for a structured verdict:

Return JSON only: {
  "verdict": "clean" | "degraded" | "catastrophic",
  "observations": string[],
  "quality_warning": string | null,
  "re_roll_suggestion": string | null
}
Check that both faces stay consistent with their references, the kiss action completes as a brief natural lip kiss, Kiss Cam text / scoreboard styling stays readable, and neither visible subject lip-syncs the off-screen PA audio.

• If verdict is clean, return the still URL and final video URL normally.
• If verdict is degraded, return the URLs plus the quality_warning so the user can review before publishing.
• If verdict is catastrophic, do not call the run complete; surface the verdict and re_roll_suggestion instead of declaring success.

Runtime expectations

• Step 1 — gpt-image-2 medium, 16:9, 2 reference images: ~40–90s
• Step 2 — kling v3-omni, pro 1080p, 15s, sound on: ~3–5 min
• Step 3 — download + emit markers: ~5–10s
• Total wall-clock per take: ~4–6 minutes

If a re-roll is needed at Step 1 the budget restarts there; at Step 2 only the video stage repeats.

Load-bearing phrases (keep verbatim)

Don't edit these without a re-validation pass — they're empirical behavior dependencies, not stylistic choices.

Image prompt (Step 1):

• spectator-POV phone screenshot ... lower bowl looking up at the giant suspended Jumbotron — without this, the model defaults to a broadcast-feed aesthetic
• iconic retro arena kiss cam style + the four decorative bullets (crimson border, sparkly hearts, star pattern, cursive Kiss Cam) — together produce the recognizable retro look
• the character from the FIRST / SECOND reference image — image-grounding lock; never replace with verbal feature descriptions
• preserve their exact likeness AND exact visual style + Do not redraw in a different style — the style-preservation lock; both halves are load-bearing

Kling prompt (Step 2):

• First frame: spectator-POV phone shot at MSG of the Jumbotron Kiss Cam segment — anchor that matches the still
• preserve their exact likeness AND exact visual style across all 15 seconds — identity + style continuity across the clip
• subtle, restrained, true-to-life motion within their original style + Neither stiff and frozen, nor over-acting + The level of expression is what a real person shows when caught unexpectedly on a stadium camera — calibrates the motion level (avoids both the "vinyl-toy frozen" and "theatrical over-acting" failure modes)
• gentle kiss on the lips — brief, sweet, natural, not staged — the kiss beat anchor (do not soften to "head" / "cheek" — the trend is a lip kiss)
• Audio (non-diegetic / OFF-SCREEN only — Subject A and Subject B stay SILENT throughout, mouths closed, do not lip-sync the dialogue below) + OFF-SCREEN gender-neutral PA announcer (NOT from either visible subject — an unseen arena voice) — without these, Kling defaults to attributing the quoted dialogue to a visible face and lip-syncs the announcer lines onto Subject A or B
• Phone mic picks up nearby fans louder than PA echo — sells the spectator-phone audio aesthetic

Kling negative_prompt (Step 2):

• exaggerated acting, theatrical expressions, over-acting, mugging at camera, cartoon reactions — without these Kling regresses toward the "expressive" failure mode where subjects mug at the lens
• subjects lip-syncing PA announcer dialogue, characters mouthing the announcer lines, subjects' mouths moving with off-screen voice, subjects talking to camera, on-screen lip-sync — paired with the off-screen audio anchor; suppresses Kling's default behavior of animating a visible mouth to match any speech on the audio track

Params:

• provider: gpt-image-2 (Step 1) — see Step 1 rationale
• prompt_adherence: strict (Step 2) — without it, scoreboard and "Kiss Cam" text drift mid-clip
• image_types: ["first_frame"] (Step 2) — pins the still as literal frame 0
• quality_mode: pro (Step 2) — 1080p output
• duration: 15 (Step 2) — the timeline beats (0-3s / 3-7s / 7-11s / 11-15s) are written for 15s; changing duration breaks the kiss-beat timing

Engine choice: gpt-image-2 + Kling-only

Step 1 — gpt-image-2, no fallback. Empirically sharpest LED panel detail (scoreboard numerals, kiss cam typography, retro decorative edges) and strongest reference-likeness lock — the combo is what sells the trend. On a moderation_blocked response, re-roll the same call rather than swapping providers while Step 1 budget remains; alternatives produced softer likeness and softer LED detail in earlier trials.

Step 2 — Kling, no Seedance. Seedance has a two-stage partner_validation_failed 422 gate (same as the baseball-trend sibling skill): an input-side gate that rejects references with recognizable real people, and an output-side gate that rejects AFTER generation if the produced clip contains recognizable-looking faces. Every Kiss Cam shot has a packed-arena crowd full of faces, so the output-side gate is unavoidable here. Kling is the only engine that lands this recipe.

Kling trade-offs: 2500-char prompt cap (recipe above is pre-trimmed to ~2400 chars; re-inflating it can trigger prompt-length errors). kling-v3-omni has no seed; identical Kling payloads can collapse to the same job/asset, so a corrective retry must materially change the first-frame still, prompt, negative_prompt, or PA audio wording. Do not submit an identical Kling payload for variation.

Failure cheat sheet

Symptom	Fix
moderation_blocked on Step 1	gpt-image-2 safety gate (often female/female subject pairings + "kiss cam" wording). Re-roll the same call within the Step 1 retry budget; do NOT swap providers — alternatives produce softer likeness and softer LED detail
Kling prompt error: prompt > 2500 chars	Re-inflated audio or aesthetic section in the Kling prompt. Cut from the audio or aesthetic block; never from the animation timeline
Scoreboard, "Kiss Cam" text, or graphics animate mid-clip	prompt_adherence not set to strict, or negative_prompt missing entries like "scoreboard changing" / "Kiss Cam text changing". Restore both params before the one Step 2 corrective retry so the payload differs
Subject identity drifts after ~8s	Reference still face crop too small — not enough facial pixels for Kling to lock. Re-roll Step 1 with a tighter face crop on the subjects only if Step 1 budget remains, then use the changed still for the Step 2 corrective retry
Subject gets redrawn in a different style (photoreal → illustrated, or vice versa)	Style-preservation lock weakened, or a specific style label (Pixar / anime / etc.) crept into either prompt. Restore the "preserve exact likeness AND visual style" anchor; remove any style label
PA announcer mispronounces a word or misses the kiss beat	Native audio is one take per Kling generation. Add a pronunciation hint, shorten the PA line, or adjust the audio timing before the one Step 2 corrective retry
One of the on-screen subjects lip-syncs / mouths the PA announcer's dialogue	Kling defaults to attributing any quoted dialogue on the audio track to a visible face in frame — without an explicit off-screen anchor, it picks a subject and animates their mouth to the words. Verify the audio block is framed as Audio (non-diegetic / OFF-SCREEN only — Subject A and Subject B stay SILENT throughout, mouths closed...) and the negative_prompt contains subjects lip-syncing PA announcer dialogue, characters mouthing the announcer lines, on-screen lip-sync; restore those anchors before the one Step 2 corrective retry
Subjects mug at camera / over-act / theatrical expressions / cartoon reactions	Animation block prescribes a loaded list of simultaneous micro-expressions, or timeline beats use exaggerated descriptors ("huge smile", "shy excited wiggle", "eyes widen"). Restore the subtle, restrained, true-to-life motion framing in the animation block; strip emotion adjectives from the timeline; verify exaggerated acting / theatrical expressions / over-acting / mugging at camera / cartoon reactions are in the negative_prompt
Kiss lands on cheek / forehead / head instead of lips	Timeline beat softened away from gentle kiss on the lips. Restore the verbatim lip-kiss line in the 7-11s beat; the trend is a lip kiss, not a peck on the head
Seedance attempted instead of Kling	Wrong engine chosen. Switch to kling — Seedance's two-stage likeness gate makes it unusable here (see "Engine choice")
Step 2 times out with no task_id returned	Client-layer timeout orphaned the upstream task — no recovery handle. Do not submit a duplicate. Surface the orphaned upstream task risk and wait for a recoverable task handle or explicit operator confirmation before any rerun

What not to do

• Don't add name chyrons. This trend has NO names — it's two anonymous subjects caught on cam.
• Don't render this as a TV broadcast cutaway with a heart overlay. It's a Jumbotron-shot.
• Don't bare-phrase "subtle motion only" (subjects freeze into vinyl-toy stiffness) OR load the animation block with simultaneous micro-expressions like "eyes widening, mouths opening, cheeks lifting" (subjects over-act and mug at the camera). The working framing is subtle, restrained, true-to-life motion at the level of someone caught unexpectedly on a stadium camera — both failure modes hide in the word "subtle"; the qualifier "true-to-life" is what calibrates it.
• Don't write timeline beats with exaggerated emotion adjectives like "huge smile", "shy excited wiggle", "eyes widen", "giggles back". Use restrained verbs: "notices", "small smile", "exchange a look", "share a quiet laugh".
• Don't name a specific target style (Pixar, anime, Disney, photoreal, etc.) in either prompt. The reference image defines the style; the prompt only says "preserve the reference style."
• Don't soften the kiss to forehead / cheek / "kiss on top of the head". The trend is a lip kiss — keep the verbatim gentle kiss on the lips — brief, sweet, natural, not staged beat.
• Don't write the audio block as bare quoted dialogue without an OFF-SCREEN anchor. Kling defaults to attributing any quoted speech to a visible face and will lip-sync the announcer lines onto Subject A or B. Always frame the audio block with Audio (non-diegetic / OFF-SCREEN only — Subject A and Subject B stay SILENT throughout, mouths closed, never lip-sync any dialogue below) and add the anti-lip-sync terms to negative_prompt.
• Don't gender either subject in the narrative or PA dialogue. Use "Subject A / Subject B", "they/them", or descriptions without pronouns.
• Don't swap providers on moderation hits — re-roll the same gpt-image-2 call only while Step 1 budget remains.
• Don't try Seedance.
• Don't generate music. Native PA announcer + crowd ambient IS the soundtrack.
• Don't run a post-processing layer (add_captions, generate_music, edit_concat, edit_text_overlay, edit_pip, any edit_*). Kling burns the scorebug + chyron + native commentary directly; anything added afterward breaks the kiss-cam illusion.