AI Skill Report Card

Detecting and Clipping Viral Moments

A89·Nov 19, 2026·Source: Web
15 / 15

Given a chunk of synchronized stream data (chat log, audio levels, STT transcript), immediately scan for convergence of signals and emit a scored clip payload:

JSON
{ "clip_id": "clip_2026_001", "stream_timestamp": "01:42:18", "duration_seconds": 38, "virality_score": 92, "trigger_signals": { "chat_spike_multiplier": "4.2x", "top_chat_keywords": ["NO WAY", "W", "CLIP THAT"], "audio_peak_db": -2.1 }, "editing_directives": { "layout": "split_screen_facecam_bottom", "hook_text_overlay": "HE ACTUALLY PULLED THIS OFF...", "caption_style": "dynamic_bold_highlight" }, "metadata": { "youtube_shorts": {}, "instagram_reels": {}, "tiktok": {} }, "publish_status": "READY_FOR_ONE_CLICK_DEPLOY" }

If no input data is provided, ask for (or simulate with placeholders) a chat log with timestamps/message rates, an audio/decibel trace, and an STT transcript — these three streams are required inputs for detection.

Recommendation▾
Add a third example showing a HELD_FOR_REVIEW case with a full scored payload (not just a near-miss) to demonstrate the gating logic more completely
15 / 15

Progress:

  • Step 1: Ingest synchronized chat, audio, and transcript streams
  • Step 2: Detect convergent trigger signals and lock T0
  • Step 3: Extract buffer window and refine clip boundaries
  • Step 4: Apply editing directives (layout, captions, zooms)
  • Step 5: Score against the 100-point Virality Index
  • Step 6: Gate — discard/queue if score < 80
  • Step 7: Generate platform-native metadata
  • Step 8: Emit structured JSON payload; stage for publish

Step 1-2: Detection

  • Compute rolling 5-minute MPS baseline; flag CHAT_SPIKE at ≥250% baseline.
  • Scan chat text for high-arousal clusters (W, L, CLIP THAT, OMG, NO WAY, emote walls).
  • Scan audio trace for decibel spikes, pitch shifts, laughter/gasp patterns.
  • Scan STT transcript for hooks: contrarian takes, punchlines, emotional peaks, strong claims.
  • Require ≥2 signals within a 15-second window to lock T0. If only 1 signal fires, log it but do not proceed to clip extraction.
  • Buffer window = T0 - 45s to T0 + 30s.

Step 3: Boundary refinement

  • Trim dead air and filler openers ("So, um...", "Basically...") from the start.
  • Snap the start point to the nearest bold statement, action frame, or curiosity gap within the buffer.
  • Ensure a setup → climax (the viral peak) → punchy outro/loop arc.
  • Final duration must land between 20–58 seconds.

Step 4: Editing directives

  • Default layout: active speaker face-tracking with auto-reframe.
  • If gameplay/screen-share is detected: split_screen_facecam_bottom (top 65% action, bottom 35% facecam).
  • Word-by-word animated captions in the safe zone (avoid bottom UI overlays), Hormozi/retention style.
  • Highlight power words/numbers/emotional keywords in yellow/cyan/green.
  • Apply 1.05x punch-in zoom at peak vocal emphasis moments only (not throughout).

Step 5-6: Virality scoring

  • Hook Strength (0–30): strength of first 2 seconds for a fast scroller.
  • Chat & Audio Intensity (0–30): magnitude of visceral audience reaction.
  • Standalone Context (0–20): comprehensibility with zero prior stream context.
  • Pacing & Retention (0–20): deduct for any dead silence > 1.2s.
  • Sum score. If ≥80, mark READY_FOR_ONE_CLICK_DEPLOY. If <80, mark HELD_FOR_REVIEW and state the lowest-scoring category as the reason.

Step 7: Metadata generation

  • YouTube Shorts: title <50 chars, curiosity-driven, include #Shorts, 2 niche tags.
  • Instagram Reels: 1-line hook caption, 3–5 targeted hashtags, call-to-comment prompt.
  • TikTok: conversational caption with SEO keyword in first sentence, 4–5 tags.
  • X: bold takeaway line, native video framing, no external links.

Step 8: Distribution

  • On [PUBLISH_NOW] or AUTO_PILOT: render spec is 1080x1920, 60fps, H.264, -14 LUFS.
  • If multiple clips queue simultaneously, stagger posts 15 minutes apart.
  • Always output the full JSON payload (schema below) before invoking any publish tool call.
Recommendation▾
Clarify how T0 and buffer windows adjust when multiple convergent signals fire close together but at slightly different timestamps
JSON
{ "clip_id": "string", "stream_timestamp": "HH:MM:SS", "duration_seconds": 20-58, "virality_score": 0-100, "trigger_signals": { "chat_spike_multiplier": "Nx", "top_chat_keywords": ["..."], "audio_peak_db": -N.N }, "editing_directives": { "layout": "face_tracking | split_screen_facecam_bottom", "hook_text_overlay": "string", "caption_style": "dynamic_bold_highlight" }, "metadata": { "youtube_shorts": {"title": "string", "tags": []}, "instagram_reels": {"caption": "string", "hashtags": []}, "tiktok": {"caption": "string"}, "x": {"caption": "string"} }, "publish_status": "READY_FOR_ONE_CLICK_DEPLOY | HELD_FOR_REVIEW" }
18 / 20

Example 1: Input: Chat MPS jumps from 8/sec baseline to 38/sec (4.75x) at stream time 00:22:10, containing repeated "NO WAY" and "CLIP THAT"; audio peak hits -1.8dB with a laughter burst; transcript shows "...I was NOT expecting that to actually work." Output: T0 locked at 00:22:10 (3 converging signals: chat spike, audio spike, semantic hook). Buffer = 00:21:25–00:22:40. Trimmed clip starts on "I was NOT expecting that to actually work" (cuts preceding filler). Layout = face_tracking (no gameplay detected). Score: Hook 27, Chat/Audio 29, Standalone 16 (needs 1 extra sentence of context, so hold start 2s earlier), Pacing 19 → 91/100 → READY_FOR_ONE_CLICK_DEPLOY.

Example 2: Input: Chat MPS rises to 180% baseline only (below 250% threshold); no audio spike; transcript contains a mild joke. Output: No T0 lock — single/weak signal only. Log as a near-miss event, do not produce a clip payload. Recommend continued monitoring.

Recommendation▾
Include guidance for handling multi-speaker streams or collab streams where facecam/speaker detection is ambiguous
  • Never lock T0 on a single signal type — always require convergence (≥2 of chat/audio/semantic) within the 15s window; this is the primary false-positive filter.
  • Prefer trimming into the buffer rather than always using the full 75s window — tightest viral clips are usually 25–40s.
  • When Standalone Context scores low, prefer pulling the start earlier (more setup) over adding text explainer cards — native context beats overlay exposition.
  • Reuse the same virality score breakdown in the output so low scores are explainable, not just a number.
  • Treat gameplay/screen-share detection as a hard switch for split-screen layout — don't apply it to pure talking-head streams.
  • Stagger multi-clip publishing even in AUTO_PILOT — simultaneous posts cannel each other's algorithmic push.
  • Don't start a clip on a hook word that lacks context — verify the viewer can understand the stakes within 3 seconds without prior stream knowledge.
  • Don't let punch-in zooms run continuously — overuse flattens the emphasis effect; reserve for true peak moments only.
  • Don't publish clips scoring ≥80 that have a Standalone Context score below 10 — low context score is a near-automatic audience-retention killer regardless of total score; flag for manual caption-card review.
  • Don't ignore dead-air penalties — a single 2s silent gap can sink an otherwise strong clip's Pacing score below threshold.
  • Don't apply split-screen layout to clips without detected gameplay/screen content — wastes vertical space and looks broken on pure facecam streams.
0
Grade AAI Skill Framework
Scorecard
Criteria Breakdown
Quick Start
15/15
Workflow
15/15
Examples
18/20
Completeness
19/20
Format
15/15
Conciseness
13/15