Toby's channel normally did about 228 views a video. Then "Taking a step back from Claude (and why you should too)" did 115,963 views in 10 days. So we are covering this channel, this video, and why it happened.
I will look at the packaging, the opening, and the rest of the edit. Let’s get started.
1. How the video performed
The video is 9:55 long and had 681 comments when this was checked. It got 115,963 views on a channel where the middle video gets 228 views, so this is not a small improvement.
The five earlier videos got 94, 228, 4,513, 219, and 361 views. The earlier high point was “Wie ich auf die langweiligste Art online Geld verdient hab” at 4,513 views, but this Claude video went far beyond that.
The video before it, “KI hat dieses Video animiert (so machst du's),” got 361 views. The other recent AI website video got 219 views, so AI as a topic alone did not make every upload work.
I read and sorted 661 viewer comments. 50.8% agree and tell their own version of the same story, while 19.5% push back by calling it clickbait or a skill issue. 11% give advice, 6.2% are praise, and only 1 comment says how the person found the video.
Also, there are no later uploads yet. So only this one worked, and that leans towards a trend or luck for now. The next uploads will tell us if Toby found something repeatable.
My verdict is that this is a real outlier, which means one video that performs far above the rest of a channel. I think the topic got picked up by YouTube’s recommendation system, but the packaging gave people a very clear reason to click.
2. What was happening around this topic
On 29 July, Anthropic reported elevated errors across all Claude models. A Claude Code creator also posted on 28 July about context waste costing people “tokens, speed, and money”, and that video got 87,421 views on a channel with 84.5K subscribers.
There was more frustration on 29 July too. Another YouTube creator said new Claude Code rules were breaking setups and wasting tokens, and it got 21,313 views on a 74.8K-subscriber channel. So this was already becoming a real topic before Toby uploaded.
On 4 August, a Reddit thread got 1,552 upvotes from a person saying, “I have usage credits off.” That is a user report, not proof of a billing problem, but it shows people were already worried about Claude costs.
Then multiple Claude models went down on 5 August. On 6 August, a Hacker News thread got 46 points and 66 comments, with one person saying, “I’ve promoted Codex to my primary harness.”
On 10 August, Riley Brown posted a video calling this period “Anthropic’s Slump”. It got 33,416 views on a channel with 276K subscribers, and compared Claude with Codex and cheaper alternatives.
The bigger wave came after Toby uploaded. A Reddit thread on 17 August got 617 upvotes and said, “4 prompts burn through my entire week.” Then another thread on 18 August got 126 upvotes and said, “Nothing except a claude status updates.”
Toby published on 11 August. The Claude problems had already started on 29 July with errors, 5 August with the outage, 6 August on Hacker News, and 10 August with Riley Brown's “Anthropic slump” video, so he came after the news started. I think that points towards a deliberate read of the mood, but we cannot see his traffic sources.
But the biggest wave, the 12 to 17 August outages and the limits anger, came after his upload and lifted a video that was already sitting there. So I would say Toby may have read the mood well, but the later outages probably gave the video more fuel than he could have planned for.
3. The creator’s other platforms
Toby has an Instagram account with 243 followers. He also has a Facebook page, but its follower count was not visible.
He runs Studio Erlach, has a newsletter, a consulting page, and an early-access product called Design Mind. But the newsletter and product user counts were not visible.
TikTok, X, LinkedIn, Threads, and a Discord or community were not found. That is a finding, not missing research.
A big outside audience would lift every video, not one. So 243 Instagram followers do not explain 115,963 views, and I personally think this came from YouTube browse and recommendations, not from Toby sending his own audience to the video.
My verdict is simple. Outside platforms did not cause this outlier.
4. The thumbnail
The thumbnail is very simple. Toby has a flat, tired expression, direct eye contact, a dark office background, giant white text saying “ruined my business”, and orange starburst shapes around his face.
Colin and Samir say, “the thumbnail has to stop your scroll,” and say a big face and symmetry work well for them in their thumbnail breakdown. This thumbnail follows that well because the face is centred, the orange shapes mirror each other, and the white and orange stand out against charcoal black.
They also say a thumbnail should “beg a question or suggest transformation.” This one creates a very direct question: what did Claude do to this person’s business?
Paddy Galloway says a thumbnail is mostly “the psychology of the click” and less about design in this video. I think that is exactly what is happening here, because “ruined my business” is not useful information by itself. It is a problem you want explained.
The thumbnail follows the short-text rule too. Film Booth says to simplify the message to around four words so it works when displayed small in this video, and Toby uses three.
But it breaks a couple of common rules. The text is not on the left, and the face is not shocked or exaggerated.
Still, that may help. MrBeast said, “I closed my mouth on all my thumbnails and the watch time went up on every video lol,” as reported by IGN. Tyler Fleischman also says a subtle expression can feel authentic in his thumbnail analysis.
The thumbnail before this used a busy editing software screen behind the face and “alles KI” text. It becomes visual noise when small, in my opinion, because the face, text, arrow, and interface are all asking for attention.
The earlier website thumbnail is cleaner, with a laptop and “AI, aber besser.” But it looks like a normal AI tool showcase, and it does not have the strange negative claim that makes you stop.
Nobody found evidence that the Adidas logo on Toby’s cap helps thumbnails. It looks incidental, which means it is just part of what he wore.
My verdict is that the thumbnail did a major part of the work. It is clean, readable, and unusual because it sells an honest confession instead of an excited AI promise.
5. The title
The title is “Taking a step back from Claude (and why you should too)”. It uses sentence case, no capital-letter shouting, no number, brackets, the Claude brand name, and a personal first-person angle.
I think the Claude name helps because it tells viewers straight away who the video is for.
Ali Abdaal says, “Use language your viewers can understand at a glance,” in his YouTube guide. This title follows that because everyone who knows Claude understands “taking a step back”.
Jake Thomas says, “The most powerful emotion when it comes to getting people to click on your YouTube title is curiosity,” in Creator Hooks’ title tests. The title gives the action, but it does not give the reason, so the viewer has to click to get it.
It also uses negative framing without saying Claude is bad. Jake Thomas says, “Negativity does a great job of getting people's attention,” and this title quietly says a popular tool may be causing a problem.
But it breaks Ali Abdaal’s mobile guideline by one character. He says titles beyond 54 characters may not show fully on some mobile devices, and this title has 55.
Also, it has no number or proof point. That means the title is clear but slightly vague, so the thumbnail has to carry the specific problem.
The bracket line helps too. Scott Nelson says, “using brackets in the title boosts my click-through rate,” in this Databox roundup.
Nobody found a credible source saying that sentence case itself improves click-through rate. Also, nobody found proof that simply naming a brand improves click-through rate, although Claude makes the audience and topic clear.
My verdict is that the title and thumbnail work as one pitch. The title says what Toby is doing, and the thumbnail says why.
6. The first 93 seconds
The intro has 16 shots. Toby’s face is on screen for 67% of it, b-roll, which is extra footage used to break up the main camera shot, takes 33%, and there are 4 text cards.
There are captions in bold white text, usually one to three words at a time. There is also music, a Claude window graphic, app icons collapsing together, a screen recording, and a meme-style car clip.
The script is:
[0:00] “I use the AI assistant Claude every single day to run my business.”
[0:05] “It single-handedly replaced almost all the other tools that I was working with.”
[0:11] “Yet, I have decided to take a step back from Claude as it started to impact my business in a negative way.”
[0:17] “And in this video, I want to tell you how things started turning for the worse and why you should consider doing the same thing.”
[0:33] “But before we dive too deep into this, I want to give you a quick overview of who I am.”
[0:43] “I'm Toby, Tobias, and I run a web design agency located in the heart of Europe in Salzburg, Austria.”
[0:55] “We started implementing it everywhere, so much that I recently upgraded to their most expensive plan.”
[1:08] “It's Britney.”
[1:17] "I'm pretty sure at a certain point it felt like cheating, like an unfair advantage."
[1:28] “Until yesterday afternoon.”
Colin and Samir say the first job of an intro is to “immediately confirm” that this is the video the viewer clicked in their intro checklist. Toby does that at 0:11, when he says Claude started affecting his business negatively.
They also say to create a new unclosed loop, which means a question the video has not answered yet. Toby does this best at 1:28 with “Until yesterday afternoon,” because now you want to know what happened.
MrBeast says, “The first minute of each video is the most important minute,” in his production handbook. Paddy Galloway also says the strongest part should come in the first 30 to 60 seconds in this Editing Podcast episode.
Toby partly follows this. The first frames confirm Claude is central to his business, the thumbnail claim comes early, and the app graphics make the point easier to understand.
But there is a problem at 0:33. He stops the story to explain who he is, and the actual incident does not begin until 1:29.
George Blackman says the hook should have only “critical information” in this video. He also says, “any more than” one credibility statement and “you're cooked,” so Toby’s business and location introduction is risky.
The intro also has a long talking-head stretch from 0:11 to 0:46. Creator Talk says a shot change every one to three seconds can keep people locked in because the brain wants novelty, in this video.
But Toby uses an off-framework move that I think works. At 1:03, he shows the Claude plan price over a red PT Cruiser clip, then says “It’s Britney” at 1:08.
Nobody found a framework source that says this specific verbal joke works. But one YouTube education channel says b-roll breaks the monotony of the main camera shot in this video, and that is what the meme does here.
The intro before this video is about AI animating the video, and its thumbnail already makes that clear. The website video before that opens around an AI website result, so both are more tool-focused and less personal.
My verdict is that the intro breaks a lot of strict framework rules, but its best move is very good. Toby turns a broad Claude opinion into a specific personal event, then stops just before telling us what it was.
7. The rest of the video
The body keeps opening loops, which are questions the viewer still wants answered. At 2:21, Toby moves into the personal failure story, at 3:29 he explains why AI became the easy option, and at 4:36 he asks what he will do about it.
He pays off the central idea at 6:26 with the 70/30 rule. That is Toby doing most of the thinking himself, then using Claude as support rather than letting it lead the work.
The video looks clean in the early section, with screen recordings, overhead desk b-roll, full-screen question cards, and dark talking-head shots. But the edit becomes much more static later.
There is a 74.7-second talking-head stretch beginning at 5:11. There is another 65-second static stretch beginning at 7:02, and a final static talking-head section near the end.
Film Booth says Shelby Church keeps storytelling and b-roll at the same pace throughout her video in this retention breakdown. Toby does the opposite, because his visual energy is high early and much lower once he starts explaining the solution.
The self-sponsor gag and subscribe ask come at 4:42. Then Design Mind gets a 50-second product section at 8:07.
MrBeast says, “BRAND DEALS ARE CONTENT,” in his production handbook. I think the earlier joke earns some attention because it is part of the tone, but the later product section is a downside because it comes after viewers have waited for the practical answer.
My verdict is that the story section is strong, but the back half gives people too many reasons to leave. The core point is useful, but it could have arrived earlier with more visual support.
8. The expected retention graph
This is a predicted retention graph from the script and edit, not YouTube’s real graph. We do not have Toby’s actual YouTube retention graph.
I think the biggest hold is the first four minutes. Toby moves from Claude helps my business to the moment where he realises he is wasting time, and that is a clear personal story with rising stakes.
The biggest predicted drop is the self-sponsor and subscribe section at 4:42. It interrupts the open question about what Toby will actually change.
The next big drop is the static explanation from 5:11 to 6:26. The useful 70/30 rule comes at the end of that stretch, but some people may leave before they hear it.
The third big drop is the Design Mind product section at 8:07. It is long enough to feel like a detour after the viewer has already received the main answer.
My verdict is that the video probably held well because of its premise and personal story, not because every section was retention-optimised. The early part is very strong, while the later part is where the edit loses momentum.
The title says the action and the thumbnail says the consequence, the promise is confirmed at 0:11, and a real problem the audience already had did the rest. That part is repeatable. But the later outages gave it more fuel than he could have planned, and there is no second upload yet, so treat it as a lever you test, not a law.
9. What you can take from this
If you are a small creator, I would copy the idea behind this thumbnail, not just the visual style. Pick a real problem, make the title explain the action, and make the thumbnail explain the consequence.
Also, show the thumbnail promise early. Toby says he is stepping back from Claude at 0:11, so viewers do not feel tricked.
But I would personally start the real story earlier. Keep your background short, and do not put the strongest incident after a long introduction.
If you already have an audience, you can take more personality risks. Toby’s flat face, deadpan joke, and personal opinion work because they feel like a person talking, not a generic AI video.
But the downside is clear too. Do not let plugs, subscribe asks, or static explanation take over the back half after you have built a strong opening.
I personally would use the same structure with a faster turn into the story and a shorter product section. But make sure you choose according to your own audience, because this video may still be a one-off spike.
10. The replicable parts, in one table
I pulled out the parts I think another creator can actually try, but I would still treat them as hypotheses.
| What this video did | Who else says it works | How sure we are, and why |
|---|---|---|
| The thumbnail carries the consequence and the title carries the action, so the two halves only make sense together. | a YouTube education channel; Paddy Galloway | Medium. The split is clear and both sources back it, but there is no CTR or impressions data and this is one video. |
| It used a flat, tired, unimpressed face instead of a shocked one. | MrBeast (via IGN); Tyler Fleischman; a YouTube education channel | Medium. Named sources back the deadpan trend, but expression changed alongside other things in one video. |
| It dropped product and UI screenshots for face plus text. | Ed Lawrence (Film Booth); a YouTube education channel | Medium. The before and after contrast is clear and sources back it, but the comparison videos had very low views. |
| It names the product, states a first-person action, and turns the story into viewer advice. | Jake Thomas, Creator Hooks; Ali Abdaal; Scott Nelson, MoneyNerd | Medium. The parenthetical appeared on earlier low-view videos too, so the product name plus negative stance looks more important. |
| It attacks a belief the target audience holds. | Aprilynne Alter and Paddy Galloway; Colin and Samir | Medium. The comments show the topic hit a live nerve, but comments show who stayed, not who clicked. |
| It says the thumbnail and title promise out loud at 0:11. | Colin and Samir; Ed Lawrence (Film Booth) | Medium. The move is clear and widely backed, but there is no real retention graph. |
| It raises the praise, then reverses on a black card before the story starts. | Hayden Hillier-Smith; Colin and Samir | Medium. The structure is visible and the sources back teasing and withholding, but the delay may have helped or merely not hurt. |
| It shows real Claude use and the plan price instead of only describing them. | MrBeast (production handbook); a YouTube education channel | Low. Show-don't-tell is backed as general advice, but nothing isolates these artefacts as a reason this video worked. |
| The outlier was the first English video after earlier German videos. | nobody we found | Medium. The language switch is a fact, but it changed with the topic, packaging, and upload gap. |
The things that looked like levers were not the levers. Fast cutting was not required, the parenthetical title had appeared on low-view videos, the intro broke several framework rules, and the cap logo was just wardrobe. The predicted retention graph and the comments cannot prove where the views came from, and this is still a one-off until another upload tests it.
My prediction is that Toby's next upload will land between 2,000 and 30,000 views by 2026-10-31, or 30 days after his next upload, whichever is later. We will score it from the next non-Short upload at 14 days old against the 228-view pre-spike median and the 115,963-view outlier. It is a hit inside that range, a miss low below 2,000, and a miss high above 30,000.
If you think I am missing something, reply to the email and let me know.