Breakout detector

Breakout Detector

Big accounts get big numbers on everything — that is baseline, not signal. A true breakout is a post that beats its own account's normal performance by a wide margin, because that gap is the market voting for a specific idea before the algorithm saturates it. This skill has no live data of its own: the user (or a paired search/scraper tool) supplies recent posts with per-account stats, and you turn that raw list into ranked, normalized breakout patterns plus a brief to make your own version fast. Speed of detection beats precision — a 70%-confidence pattern acted on this week beats a perfect one found after saturation.

1. Gather and normalize

1. Collect 30-100 recent posts (last 7-30 days) across 10+ accounts in the niche. For each, require: account follower count, that account's typical/median views, this post's views, and likes+comments+shares+saves.
2. Compute the Outlier Score two ways and keep the higher:
   • View Multiple = post views ÷ that account's median views (best signal; needs per-account baseline).
   • Reach Ratio = post views ÷ follower count (fallback when you lack an account baseline).
3. Flag as breakout candidate if View Multiple ≥ 3x OR Reach Ratio ≥ 5x. Discard anything under 2x — it's baseline.
4. Kill false positives: drop posts inflated by paid ads, a collab with a far-bigger account, or a one-off news spike unrelated to the niche.

2. Isolate the driver

For each candidate, name the ONE variable most responsible. Score every candidate across:

1. Format — carousel, talking-head, green-screen, listicle, B-roll voiceover, text-on-screen.
2. Angle/hook — contrarian take, "I was wrong about X", before/after, mistake-confession, us-vs-them.
3. Sound/audio — trending sound ID, original VO, specific song. Note if the same sound repeats across candidates.
4. Structure — hook <2s, payoff timing, loop, open-loop CTA. Cluster candidates that share a driver. A pattern only counts when 3+ different accounts hit outlier scores using the same driver — one viral post is luck, three is a pattern.

3. Score breakout potential

Rate each pattern 1-5 on three axes, then multiply:

• Novelty (is it still early, few imitators?) × Proven demand (how many accounts + how high the multiples?) × Your execution fit (can YOU ship this in <48h with your skills/assets?). Score of 45+ (of 125) = make it now. Below 20 = skip; either saturated or off-fit.

Output template

NICHE: <niche> | Window: <dates> | Posts analyzed: <n>

TOP BREAKOUT PATTERNS (ranked)
1. <Pattern name> — Driver: <format/angle/sound>
   Evidence: <n> accounts, avg <Xx> outlier score, e.g. @acct 8.2x
   Score: Novelty _/5 × Demand _/5 × Fit _/5 = __/125
   Why it's spiking: <1 line, first-principles>

MAKE-YOUR-VERSION BRIEF (pattern #1)
- Hook (first 2s): <copy-ready line>
- Format/shot list: <specifics>
- Sound/audio: <ID or type>
- Your angle twist: <how you make it yours, not a clone>
- Ship-by: <date, within 48h>

SKIP LIST: <saturated or off-fit patterns + why>

Platform variants

• TikTok/Reels/Shorts: View Multiple is king; watch sounds — a shared trending sound across candidates is the strongest early signal. Saturation window is 3-10 days.
• YouTube: use views ÷ channel median and outlier thumbnails/titles; breakouts live longer (weeks), so novelty decays slower.
• X/Threads: use impressions ÷ follower count; driver is usually angle/hook, not format. Decay is hours.
• LinkedIn/Instagram feed: saves and shares outweigh likes; normalize on those.

Rules

• Never rank by raw views or follower count — always normalize to the account's own baseline, or you'll just rediscover big accounts.
• A pattern needs 3+ independent accounts hitting outlier scores; one hit is noise.
• Detect the driver, not the topic — copying the topic saturates; copying the format/angle transfers.
• Act inside the saturation window (days, not weeks) or the edge is gone — favor speed over a perfect read.
• Always add a twist so your version isn't a clone; the algorithm punishes duplicates.
• Demand you get fresh input data each run — this skill scores what it's fed and never claims live knowledge.