--- name: reverse-engineer-niche description: Use when the user wants to reverse engineer a YouTube niche, or a channel's competitors, into a guidelines report of its outliers, winning titles, thumbnails, hooks and scripts, the versions that flop, and video ideas. Requires the TubeLab MCP. license: MIT metadata: author: tubelab version: "4.1" requires: tubelab-mcp --- # Reverse Engineer Niche ## Overview A niche taken apart through 12 to 20 channels, emerging ones first, as a niche guidelines book. Five lenses: - **Working formats:** a new video must fit the format and length viewers already watch. - **New and specific:** new enough to stop a scroll, specific enough to feel real: a price, a job, a place. - **Anchors:** names, brands, tools or terms that tell YouTube which viewers to send. - **Dead versions:** a winner's promise that flopped. - **Who can copy it:** anyone, or only someone with its asset: experience, trust, credentials, a skill, fame or budget. Not for last month's winners (`trend-report`), one channel (`reverse-engineer-channel`) or whether to enter (`niche-validator`). **Without TubeLab:** no search for new or similar channels, no audience, income, lost-ads or terminated-channel data. The report covers only channels you already know. ## Workflow 1. Ask once, with any scope question: end with video ideas, for their channel or a new one here? 2. Run the searches below, up to 15, never paging one. 3. Verify each channel with a free read (it gives the avatar): on topic, long videos, the user's language. English titles can hide Hindi audio; the hit's audience region shows it. Keep 12 to 20. 4. Median each channel's last three months of long videos. Winners: its top three at twice that or more. Flops: under half, older than a month, trailing their neighbours. One set of totals everywhere. 5. Fill the audience and income page from channel hits only; video hits carry just ads and RPM. 6. View every winner's and every flop's thumbnail on contact sheets. 7. Group winners into three to five patterns, each on three or more channels, each winner explained against its channel's normal videos. 8. Count title formulas and thumbnail compositions across winners and flops. Rising: also in the trending search. 9. Read the opening 30 seconds of ten or more winners on different channels. A hook formula is a beat order, named to fit any video, shared by two or more channels. 10. Read four to six full winning transcripts from different patterns. Mark each beat and its time. A video formula is a layout two or more share. 11. Pair dead versions with their winners, same channel when you can, with the reason. 12. Count topics, sub-topics and power words: winning titles against flops. Worn out: as common in flops. 13. Run the risk searches and fill the risk page. 14. Blend two patterns into a new angle no sampled video has filmed. 15. Given a channel, find up to three gaps: patterns its competitors win with that it skips. 16. If wanted, run `video-ideation` on these findings for the last chapter. Unanswered: offer it in the closing line. ```jsonc // One search per line. {niche}: one to three phrasings of one niche. {code}: the user's language. // Keep publishedAtFrom and classificationQuality as written. Only the risk searches leave quality out. // 10 searches as written. Spend the spare five on a second phrasing, a wider date range, or handle lookups // ({ "query": ["@handle"] }, nothing else) for big sampled channels found only by video, so the audience page covers them. // Channels created this year, then emerging channels. { "query": ["{niche}"], "language": ["{code}"], "publishedAtFrom": "trimester", "classificationQuality": "neutral", "joinedDateFrom": "year", "medianViewsFrom": 1000 } { "query": ["{niche}"], "language": ["{code}"], "publishedAtFrom": "trimester", "classificationQuality": "neutral", "subscribersFrom": 3000, "subscribersTo": 300000, "avgViewsToSubscribersRatioFrom": 1 } // Outliers: small channels' breakouts, breakouts on channels with 30 videos or fewer, then established channels' breakouts. { "query": ["{niche}"], "language": ["{code}"], "type": "video", "publishedAtFrom": "trimester", "classificationQuality": "neutral", "averageViewsRatioFrom": 2, "subscribersCountFrom": 3000, "subscribersCountTo": 300000, "size": 40 } { "query": ["{niche}"], "language": ["{code}"], "type": "video", "publishedAtFrom": "trimester", "classificationQuality": "neutral", "averageViewsRatioFrom": 2, "channelVideoCountTo": 30, "subscribersCountFrom": 1000, "size": 40 } { "query": ["{niche}"], "language": ["{code}"], "type": "video", "publishedAtFrom": "trimester", "classificationQuality": "neutral", "averageViewsRatioFrom": 1.5, "subscribersCountFrom": 300000, "size": 40 } // Similar channels, one seed per search: the best breakout, or the given channel. Several seeds blur into a generic set, // and the hits are mostly small, recently found channels: established ones come from the outlier search above. { "relatedChannelId": [""], "language": ["{code}"], "publishedAtFrom": "trimester", "classificationQuality": "neutral", "subscribersFrom": 3000, "size": 40 } // Outliers of similar channels, same seed. { "relatedChannelId": [""], "type": "video", "publishedAtFrom": "trimester", "classificationQuality": "neutral", "averageViewsRatioFrom": 2, "size": 40 } // Trending title formulas whose use grew 10% or more over the previous period. English only: pass no language. // Keep only formulas whose example titles fit the niche; the rest are generic. { "query": ["{niche}"], "lastVideosCountChangePercentageFrom": 10 } // Risk, seeded with the three strongest emerging channels (live seeds work): similar channels terminated, then that lost ads. { "relatedChannelId": ["", "", ""], "terminated": true, "publishedAtFrom": "year", "size": 40 } { "relatedChannelId": ["", "", ""], "monetizationChanged": true, "monetizationAdsense": false, "publishedAtFrom": "year", "size": 40 } // AI-made channels: count detections.madeWithAI on the channel hits above. A search filtered on it returns AI films, off topic. ``` ## Output Format One HTML book from this template, style and script unchanged. Repeat or drop page types as the data allows. ```html {Niche}: Niche Guidelines {CARD:
{ratio}×{title}{channel} · {views} views · {date}
{ratio}: views ÷ the channel's three-month median from step 4, the number that picks winners and flops, on every badge and ratio. Never print TubeLab's averageViewsRatio. ROW CARD: the same with class="v h", everything after the image link in one
, and

{why, one line}

last when the page asks for it. FLOP: class="x" on the image link crosses it out. Any other ratio, outside a card: {ratio}×, class="r lo" for a flop's. PAGE:

{Chapter · section}

{title}

{body}
. Each body below is one page. A page that overflows becomes two pages.}
{16 winners: }

{Competitors of {channel} · }{run date}

{Niche}:
Niche Guidelines

What wins, what flops and how to make it, from {n} winners and {n} flops.

{n} channels: {n} emerging and {n} established · {n} videos cited

01

Research

The niche · Working formats · The channels · Who watches, and how it pays · Outliers · What to avoid · What got similar channels in trouble · Patterns that work

{The niche:

{what it is}

Who watches
{the most common age, gender, income tier and region across the channel hits, as an estimate}
What wins
{with numbers}
Start with
{the pattern to make first, and why}
{CARD}
} {Working formats: the same split. kv: Format, Length, Faceless or on camera ({n} of {n} winners are on faceless channels, from the hits and checked on the thumbnails), Where to differ. Right:
{four CARDs}
} {The channels, h2 "{n} channels: {n} emerging, {n} established":
{channel}{emerging or established} · {subs} subscribers
median {views} views · {ratio} per subscriber

Medians cover each channel's last three months of long videos. Left out: {channels and why}.

} {Who watches and how it pays, h2 "Who watches, and how it pays":

Who watches

Age
{most common band} · {n} of {n} channels
Gender
{male, female or balanced} · {n} of {n}
Income
{low, middle, upper-middle or high} · {n} of {n}
Region
{region} · {n} of {n}

How it pays

Ads on
{n} of {n} channels
RPM
{band}
Also earn from
{one per detected method (sponsors, affiliate links, products, courses, newsletter, community, services, donations):
{method} · {n} channels
}

Estimates from TubeLab's channel search, for {n} of the {n} channels. No estimate for: {channels found only through a video search}. {If the fields come back empty: "not available on this plan".}

} {Outliers, two pages of six:
{ROW CARDs with why}

{n} of {n} are also outliers against their channel's whole list, by TubeLab's flag. The rest are recent wins only: {titles}.

} {What to avoid:

{what the flops share}

{six FLOP ROW CARDs with why}
} {What got similar channels in trouble, from the risk searches. Keep only channels that share a sampled title formula or thumbnail style. If none is close, one line saying so:

{n} similar channels were terminated and {n} lost ads in the last year. {what their formats share}

{one ROW CARD per channel, up to six, for the video it is cited by: {terminated | lost ads {Mon YYYY}} in place of the ratio, a terminated channel's image from https://cdn.tubelab.net/vi/{id}.jpeg, with the channel only, and

{its title formula, slots in [ ], and its look}

last. Four fit under the lead; the rest and the note go on a page headed "What got similar channels in trouble, continued"}

The data shows what happened, never why. AI-made channels found in the niche: {n}, {n} of them classed low effort. They share: {format}.

Never print this note: a terminated channel's thumbnails load from https://cdn.tubelab.net/vi/{id}.jpeg.} {One page per pattern, h2 = its name:

{n} winners on {n} channels

{Angle, New and specific, Title (formula, slots in [ ]), Thumbnail (composition name), Hook (formula name), Script (video formula name, only when read), Who can copy}
{three ROW CARDs}
}

02

Packaging

Anatomy of an outlier · Title formulas · Winning words · Thumbnail design

{Anatomy of an outlier:
{ANNOTATED}

{its title, the slots in }

{channel · views · date · ratio}

{In the title, In the picture, Always}
} {ANNOTATED, boxing what matters on one real thumbnail (face, payout screen, amount, arrow), in percent of the image:
  1. {what box 1 holds, words on it}
} {One page per title formula, no h2:

"{formula, slots in [ ]}"

{n} winners on {n} channels

{n} flops

{Rising across YouTube: {n} videos use it, {ratio}× on average, use up 10% or more · click triggers: {triggers} | Seen in this niche only}

{three ROW CARDs}
} {More formulas, four a page:

"{formula}"

{counts}

{ROW CARD}
} {Worn out, two a page:

"{formula}"

{counts}

{winner ROW CARD}{FLOP ROW CARD}
} {Winning words:

Win

{CHIPs}

Worn out: as common in flops

{CHIPs with class="chip worn"}
CHIP:
{word or topic} · {n} winners, {n} channels, {n} flops
e.g. {title} · {ratio}×
} {One page per thumbnail composition, h2 = its name:

{n} winners on {n} channels · {n} flops

{ANNOTATED, legend moved right}{proof}
{legend}
Text and colour
..
Why it likely works
..
{two CARDs}
}

03

Storytelling

Hook formulas · Video formulas

{One page per hook formula, h2 = its name:

{n} of {n} openings read, on {n} channels · the first {n} seconds · Who can copy: {..}

{beat}{beat}
{title} · {channel · views · date · ratio}

{Beat}"{its words}"

} {One page per video formula, h2 = its name:

{n} of {n} scripts read

{Beat: Hook, Setup, Re-hook, Stakes increase, Peak, Twist or Payoff} {what happens}
0%50%100% of the runtime
{Main payoff, What repeats, Example: linked title, its stats and length}
}

04

Learnings

Do's and don'ts · New angles{ · Gaps to win}

{One page per pair, h2 = the lesson:
✓ Do{ratio}×
{title}{channel · views · date}
{the dead version, the same with ✕ Don't and class="x" on its image link}

{the reason}

} {One page per new angle, two patterns blended, h2 = the angle:

{pattern}

{CARD}

+ {pattern}

{CARD}
{New angle, Why it should work, Filmed yet? (not in the {n} videos counted)}
} {Given a channel: one page per gap, h2 = the gap:
{two competitor CARDs}
{Their proof, Why it fits}
. Then a page "Overall": kv of Copy, Avoid, Position, each .}
{Only after video-ideation ran.}

05

Video ideas

{how many ideas, for whom, from what}

{Overview:
#IdeaFromPotentialFit
} {One page per idea, h2 = "{title}". Each takes a title formula, composition and hook from this report:
{From, Title formula, Thumbnail (composition: its parts for this video), Hook (formula: its beats for this video), Avoids (linked dead version, ratio, why), Potential · Fit}

Proof

{CARD}
}

Next steps

What to run next

{skill} {what to do}

"{a prompt to paste, [brackets] for what the user fills in}"

Made with the YouTube research skills on the TubeLab MCP · Install: npx skills add TubeLabHQ/youtube-research-skills

``` - File: `reverse-engineer-niche-{niche-slug or handle}.html`, or an artifact. Chat: the top finding and the file. - Plain, short language, with causes stated as likely. One idea and a few lines per page. - Link every video, channel and moment from the data. - Spell formulas and compositions out where used, never by page number. ## Common Pitfalls - Do not call one channel's hits a pattern. One viral video proves the idea, not the channel. - Do not sell a pattern seen only on established channels to a new channel. It may ride on trust. - Do not say why a channel was terminated or lost ads. The data does not show why.