Find Outlier Videos in Your Niche with Claude
An outlier is a video that got far more views than its own channel usually gets. Have Claude pull the last 90 days of videos from 10 to 20 channels in your niche with Apify, score each video as views ÷ that channel's median views, and keep everything at 3× or more. Then have Claude read the first 30 seconds of each outlier's transcript and write an idea bank of hooks, title patterns and formats. Twenty channels cost roughly $2 to $4 in Apify usage per run.
Sorting a niche by raw views only tells you which channels are big. A 500,000-view video is a flop for a channel that averages two million and a breakout for one that averages 50,000. The second one is the useful signal: the topic, title or hook did something the channel's audience size can't explain. This guide finds those videos, then uses the transcript to see what the first 30 seconds actually said.
Paste this and let Claude set it up
If you have Claude Cowork connected to Apify, paste the prompt below. Claude asks you a few questions, runs a two-channel test and shows the cost, then builds the outlier table and the idea bank. It asks before every step that spends money or writes a file. The rest of the post explains each step so you can check its work or do it by hand.
Claude interviews you, tests on two channels and shows the cost, then scores every video against its channel's median and writes an idea bank from the outliers' first 30 seconds.
- Asks for your niche, seed channels, platforms, window, threshold and folder, one question at a time
- Checks the Apify connector and reads the three Actors' current prices
- Runs a two-channel test and shows the cost and how far back the data reaches
- Scrapes every channel, scores each video and flags paid or suspicious spikes
- Pulls the YouTube outliers' transcripts and writes the idea bank to your folder
Help me find outlier videos in my niche and turn them into a bank of video
ideas. Work step by step and ask me before anything that costs money or
writes files.
1. Ask me these questions, ONE at a time, waiting for each answer:
a. My niche, and what my own channel makes (one or two sentences).
b. Seed channels: YouTube channel URLs and/or TikTok handles, up to 20.
Tell me to copy handles from each creator's own profile links, not to
guess them.
c. Which platforms to include: YouTube long-form, TikTok, or both.
d. The window (default: the last 90 days).
e. The outlier threshold (default: 3x the channel's median views).
f. Which folder to save the results in.
2. Check that the Apify connector is available by looking for its tools
(search-actors, fetch-actor-details, call-actor, get-dataset-items). If
they are missing, stop and tell me how to add it: Customize > Connectors >
Add custom connector, URL https://mcp.apify.com, then sign in to Apify.
3. Use fetch-actor-details on streamers/youtube-scraper,
clockworks/tiktok-profile-scraper and codepoetry/youtube-transcript-ai-scraper
and tell me their current price per result for my plan.
4. Test run (ask me first). Pick one YouTube channel and one TikTok handle
from my list (or two of one platform).
- YouTube: call streamers/youtube-scraper with startUrls = [{"url": <channel>}],
maxResults = 15, maxResultsShorts = 0, maxResultStreams = 0,
sortVideosBy = "NEWEST".
- TikTok: call clockworks/tiktok-profile-scraper with profiles = [<handle>],
resultsPerPage = 15, profileSorting = "latest", excludePinnedPosts = true.
Read results with get-dataset-items. Report per channel: videos returned,
the oldest publish date returned, any error or empty result, and the cost.
If the oldest date is newer than the start of my window, tell me how many
results per channel I need instead. If a handle returns nothing or only
old posts, tell me it is probably the wrong account.
5. Full run (ask me first, with an estimated cost): same inputs for every
channel, with maxResults / resultsPerPage set high enough to cover my
window.
6. Score the videos. Treat each channel AND platform separately (a creator's
YouTube and TikTok are two baselines):
- Keep videos published inside my window.
- Mark videos younger than 7 days as "too new" and leave them out of the
median and the ranking.
- If a channel has fewer than 5 eligible videos, say so and do not rank it.
- median = the median views of the channel's eligible videos.
- outlier score = video views / median, to two decimals.
- like rate = likes / views.
- Flag a video "check before copying" if its caption or title marks it as
an ad or paid partnership (#ad, "sponsored", "paid partnership"), or if
its like rate is below a third of the channel's median like rate.
Show me a table of every video at or above my threshold: platform,
channel, title or caption start, publish date, views, channel median,
score, like rate, flag, URL.
7. Transcripts (ask me first, with the cost): call
codepoetry/youtube-transcript-ai-scraper with startUrls = the YouTube
outlier URLs, languages = ["en"] (or my language), outputFormats = ["json"]
and enableAiFallback = false. From transcript_json, take every segment
that starts in the first 30 seconds. For TikTok outliers, use the caption
text and the duration instead.
8. Write the idea bank:
- For each outlier: what the first 30 seconds promise (paraphrase, plus at
most one short quoted line), the title pattern, the format and length,
and the topic angle. Say plainly when the opening does not explain the
result (for example, it opens with an announcement).
- Patterns that repeat across two or more outliers.
- 10 video ideas for MY channel that reuse a pattern, not a title. Never
copy another creator's title or script.
Save it as outlier-ideas-<YYYY-MM-DD>.md in my folder (ask me first), and
save the scored table as outliers-<YYYY-MM-DD>.csv next to it. Then tell me
the total Apify cost.
Claude asks before each change it makes. Read each step before you approve it.
What counts as an outlier
The definition matters more than the tooling. This guide uses one formula and three rules.
Outlier score = a video's views ÷ the median views of its own channel's eligible videos in the window
- The window is the last 90 days. It is long enough to give most channels 5 or more videos, and short enough that the channel's audience size hasn't changed much.
- Eligible means published inside the window and at least 7 days old. A 4-day-old video hasn't finished collecting views, so including it drags the median down and its own score looks worse than it is.
- Each channel and platform is its own baseline. A creator's TikTok and YouTube audiences are different sizes, and YouTube Shorts behave differently from long-form videos. This guide scores long-form YouTube and TikTok separately and leaves Shorts out.
- The threshold is 3×. Anything at 3× or above goes into the idea bank. Between 1.5× and 3× is worth a glance, not a pattern.
Why the median and not the average: one viral video pulls an average up and hides the next outlier. The median ignores it.
A channel needs at least 5 eligible videos before its median means anything. Below that, one good week moves the baseline too much, so the setup prompt reports the channel instead of ranking it.
What a real run found
We ran this on 2026-10-21 in the home-coffee niche: three YouTube channels at 15 newest long-form videos each, and one TikTok profile at 15 newest posts. Here is every video that scored 3× or more.
Real output from a test run on 2026-10-21. Views as scraped that day; window 2026-07-23 to 2026-10-21; videos under 7 days old excluded from the median.
| Platform | Channel | Video | Published | Views | Channel median | Score | Like rate |
|---|---|---|---|---|---|---|---|
| TikTok | morgandrinkscoffee | "#ad an elevated Creamsicle-Inspired coffee drink…" | 2026-09-01 | 7,700,000 | 281,700 | 27.33× | 0.3% ⚠ |
| TikTok | morgandrinkscoffee | "Do not put milk in your kettle, people. Pop it in a jar with your ground coffee." | 2026-08-26 | 2,100,000 | 281,700 | 7.45× | 9.2% |
| YouTube | Lance Hedrick | This Will Change How You Store Coffee | 2026-09-05 | 767,994 | 174,533 | 4.40× | 1.7% |
| YouTube | Morgan Eckroth | We're Done With Pumpkin Spice Lattes. Make This Instead. | 2026-09-29 | 346,584 | 97,780 | 3.54× | 2.0% |
| YouTube | Morgan Eckroth | How To Make EASY And Delicious Milk-Brewed Coffee At Home | 2026-09-18 | 304,709 | 97,780 | 3.12× | 1.8% |
Four things the table shows that a "most viewed" list would not:
- The biggest number is the least useful. The top TikTok scored 27× but is marked
#ad, and only 0.3% of viewers liked it, against a channel median of about 12%. That pattern fits paid distribution. It tells you what a brand budget can do, not what the audience wanted. This is why the prompt flags ads and low like rates as "check before copying". - The same idea broke out twice. The milk-in-a-jar TikTok (26 August) and the milk-brew YouTube video (18 September) are the same technique. In the YouTube video's first 30 seconds, the creator says she made "a short form video about milk brew a couple of weeks ago" and that this video goes deeper. A short-form outlier followed by a long-form deep dive three weeks later is a pattern you can reuse.
- The raw-views leader didn't make the list. James Hoffmann's channel had the most views per video by far, but nothing in the window reached 3× his median (his best was 1.55×). His channel also had only 5 videos inside the window, close to the minimum.
- Different channels, different baselines. Lance Hedrick's 768,000-view video scored 4.40×. The same view count on James Hoffmann's channel would have scored about 1.4×.
Why the transcript matters: the first 30 seconds
A title tells you the promise. The opening tells you how the video kept people watching. We pulled the three YouTube outliers' transcripts with the YouTube Transcript Scraper Pro and kept every caption segment that started in the first 30 seconds. All three had auto-generated captions, so no AI transcription was needed. (How that Actor works, in full.)
Real output from a test run on 2026-10-21, paraphrased with one short quote each.
| Outlier | What the first 30 seconds do |
|---|---|
| This Will Change How You Store Coffee (4.40×) | Opens on the viewer's own moment, "You just bought a great bag of coffee. What do we do now?", lists four storage options the viewer is probably weighing, then promises a conclusion drawn from the recent scientific literature. Problem, options, verdict, all inside 30 seconds. |
| How To Make EASY And Delicious Milk-Brewed Coffee At Home (3.12×) | Names the method, then states the payoff: something close to a latte without touching an espresso machine. It explicitly builds on an earlier short-form video and promises more ways to do it. |
| We're Done With Pumpkin Spice Lattes. Make This Instead. (3.54×) | Opens with a book announcement and a thank-you, not the drink. The opening doesn't explain this result; the seasonal title, published at the start of pumpkin-spice season, is the more likely driver. |
The third row is the honest limit of the method. Sometimes the opening is ordinary and the timing or the title carried the video. The setup prompt tells Claude to say so instead of inventing a hook.
From these three, the idea bank's patterns were:
- "You just bought X. Now what?" Start at the moment right after a purchase, then give a verdict.
- Latte results without the expensive machine. An accessibility promise in the first sentence.
- Seasonal swap titles. "We're done with X. Make this instead," published as the season starts.
- Test short, go long. A short video that breaks out becomes the long-form topic a few weeks later.
The manual workflow, step by step
The setup prompt does all of this for you. Here is the same process in plain steps, with the inputs that worked in the test.
1. Pick seed channels that are actually comparable
Choose 10 to 20 channels that make the kind of video you make, at roughly your size or up to 10 times larger. Copy each handle from the creator's own website or profile links. In our test, two TikTok handles we guessed from the creators' YouTube names returned an empty account and an account with no posts since 2024. A wrong handle doesn't fail loudly; it just quietly gives you the wrong baseline.
2. Pull YouTube videos
Use the YouTube Scraper (streamers/youtube-scraper), with one entry in startUrls per channel:
| Input | Value | Why |
|---|---|---|
startUrls | [{"url": "https://www.youtube.com/@channel"}] | One entry per channel |
maxResults | 15 to 30 | Enough to cover the window |
maxResultsShorts / maxResultStreams | 0 | Keep Shorts and streams out of the long-form baseline |
sortVideosBy | NEWEST | Newest first |
Each video comes back with viewCount, likes, date, duration, title and url. Check the oldest date each channel returned: if it's newer than the start of your window, raise maxResults for that channel. In our test, 15 videos covered almost exactly 90 days for a channel posting more than once a week, and 11 months for one posting less often. The Actor also has a date-filter input, oldestPostDate, but it's a paid add-on per result, so trimming by date in Claude is cheaper.
3. Pull TikTok posts
Use the TikTok Profile Scraper (clockworks/tiktok-profile-scraper) with profiles (handles without the @), resultsPerPage, profileSorting: "latest" and excludePinnedPosts: true. Pinned posts are usually old hits and would distort a 90-day baseline. Each post returns playCount, diggCount (likes), createTimeISO, the caption in text, and videoMeta.duration.
4. Score
Apply the formula and the rules above, per channel and per platform. Claude does this reliably if you give it the exact rules; the setup prompt's step 6 is the version we tested against the real data.
5. Read the openings
Send the YouTube outlier URLs to codepoetry/youtube-transcript-ai-scraper with outputFormats: ["json"]. Each result has a transcript_json list of segments with start times in seconds, so "the first 30 seconds" is every segment with start under 30. Keep enableAiFallback off unless a video has no captions, because AI transcription is billed per minute.
For TikTok, the caption and duration are usually enough. The TikTok Actor can also download subtitles or transcribe videos, but transcription is a paid add-on per minute.
6. Write the idea bank
Ask Claude for patterns that repeat across at least two outliers, and for ideas that reuse a pattern rather than a title. "You just bought X, now what?" is a pattern you can use for any product in your niche. Another creator's exact title is not yours to reuse.
What it costs
Prices are per result on the Actors' pages, checked on 2026-10-21. Paid Apify plans pay less per result than the free plan.
| Step | Actor | Free plan | Starter plan |
|---|---|---|---|
| YouTube videos | streamers/youtube-scraper | $0.004/video subject to change · verified 2026-10-21 | $0.003/video |
| TikTok posts | clockworks/tiktok-profile-scraper | $0.003/post subject to change · verified 2026-10-21 | $0.002/post |
| Transcripts | codepoetry/youtube-transcript-ai-scraper | $0.001/transcript subject to change · verified 2026-10-21 plus $0.0025 per run | $0.0009/transcript plus $0.00225 per run |
Our test returned 45 YouTube videos plus one "channel does not exist" row for a mistyped handle, and 15 TikTok posts. The 46 YouTube results billed $0.138 and the TikTok posts $0.03 on a Starter plan, exactly the listed rates; the error row was billed like a video. The two wrong TikTok handles cost another $0.022, which is the price of guessing. Three transcripts at the listed Starter rate come to under half a cent.
A full monthly run of 20 YouTube channels at 25 videos each is 500 videos: about $2.00 on the free plan or $1.50 on Starter. Adding 20 TikTok profiles at 30 posts is 600 posts: about $1.80 or $1.20. Ten transcripts add about a cent. Claude's own usage comes from your Claude plan, not Apify.
Limits worth knowing
- Views are a snapshot. Run it on a different day and the numbers move, especially for videos only a few weeks old. The 7-day rule reduces this but doesn't remove it.
- Paid reach looks like a breakout. An ad or boosted post can top the table. The like-rate check catches the obvious cases, not all of them.
- The opening doesn't always explain the result. Timing, a thumbnail or an outside link can carry a video. Claude can't see thumbnails in this workflow.
- Small channels have noisy medians. Five eligible videos is a floor, not a comfortable sample.
- Ideas, not copies. Use the patterns. Don't lift titles, scripts or thumbnails from the creators you studied.
- Platform terms. This reads public video pages and metadata through Apify's Actors. Check YouTube's and TikTok's terms for your use, and see whether scraping is legal for the general picture.
FAQ
A video that got far more views than its own channel usually gets. This guide scores it as the video's views divided by the median views of the channel's videos from the last 90 days, and treats 3x or more as an outlier.
One viral video pulls the average up and makes the next outlier look normal. The median ignores a single extreme video, so it is a steadier baseline.
Yes, but score Shorts separately from long-form videos. They get very different view counts, so mixing them in one median hides outliers in both. Set maxResultsShorts in the YouTube Scraper to collect them.
About $2 on Apify's free plan for 500 YouTube videos, or $1.50 on Starter, plus about $1.20 to $1.80 if you add 600 TikTok posts. Transcripts for ten outliers cost about a cent. Prices as listed on 2026-10-21.
Yes. In Claude Cowork you can turn the final prompt into a scheduled task. See our guide to Cowork scheduled tasks for the setup.
Start with five channels
You don't need 20 channels to see whether this works for your niche. Paste the setup prompt, give Claude five channels you already watch, and look at the outlier table before spending more. If nothing clears 3×, lower the threshold to 2× or widen the window before you widen the channel list. You'll need an Apify account; the free plan includes $5 of platform usage a month, enough for a five-channel test several times over. For more creator prompts, see Claude prompts that use live web data.