Ten Tutorials In, One Confident Agent Out. Nobody Checked If They Were Right.
This one is not a list of tools. It is a build guide, and a genuinely good one – four numbered steps that take you from “my coding agent cannot watch YouTube” to “my coding agent is an expert editor”.
01 The gap. Paste a YouTube link into Claude Code and it gets a URL. No frames, no audio. 02 Give it ears. A free SKILL.md from GitHub pulls the transcript – every word said. 03 Give it eyes. A free key from Google AI Studio plus the YT Analysis MCP, and now it has the actual frames and the sound. 04 Build expert agents. Feed it the ten best videos on a skill; each video becomes a skill; out comes an editor agent.
Every one of those steps is real and the order is right. I want to add one step that is not on the card, and flag one prerequisite that is on the card but not labelled as what it actually is.
The missing step is 05: verify. Because “ten editing tutorials in, editor agent out” is a claim about knowledge, and nothing anywhere in that pipeline asks whether the ten tutorials were correct.
The card this post is about
Watch the original reel: “Create any Claude Code agents from YouTube”
Four steps – the gap, ears, eyes, expert agents – footed “Tools and features change often. Check each official source before you build.”
Ears first, then eyes – and the card is right about the order
The line under step 01 is the most useful sentence on the whole card: “Give it ears first, then eyes.” That is not stylistic. It is a cost decision, and almost everybody gets it backwards.
A transcript is text. It is small, cheap, searchable, and you can feed a hundred of them to an agent without thinking about it. Frames are the opposite: heavy, expensive in tokens and quota, slow, and you usually only need a handful of them.
So the question before you turn on the eyes is not “do I want them” – of course you want them – it is where does the knowledge in this video actually live: in the words, or in the pixels?
Diagram: does this video need ears, or eyes?
| What the tutorial is teaching | Sounds like | Where the knowledge lives | What you need | Rough cost |
|---|---|---|---|---|
| A concept or a reason | “The cut works because the viewer is still finishing the last sentence” | The words. The picture is illustration | Ears only – the transcript | Almost nothing |
| A sequence of named steps | “Open settings, then export, then choose H.264” | The words, as long as the names are spoken | Ears only | Almost nothing |
| A position on screen | “Click the little thing up here”, “drag it over there” | The pixels. The transcript is useless – it literally says “here” | Eyes – frames at those timestamps | Real tokens and quota |
| A visual result or comparison | “See the difference?” … silence | The pixels, and the judgement is yours | Eyes, plus a human looking at it | Real tokens, plus your time |
| Timing and rhythm | A demonstration with no narration at all | The pixels and the audio together | Eyes and ears, same pass | The most expensive case |
Row three is the one that converts people. Any tutorial where the presenter says “click this” is a tutorial whose transcript is actively misleading – it reads as instructions and contains none. That is exactly the gap the MCP fills, and it is also the reason you should not turn the eyes on for everything: four videos out of ten probably do not need them at all.
The four steps, as printed, plus one
Column 2 below is the card’s own text. Columns 3 and 4 are ours – what each step quietly requires, and what to watch.
| Step | What the card says | What it actually asks of you | Our note (not on the card) |
|---|---|---|---|
| 01 The gap |
“Claude Code can’t watch videos. Paste a link and it gets nothing.” What it sees: just a URL. No frames, no audio. The fix: give it ears first, then eyes. |
Nothing. This step is a diagnosis | Accurate. A link is a string. Nothing about it carries the video |
| 02 Give it ears |
“Install the YouTube skill and it hears every word of any video.” Source: free SKILL.md on GitHub. What it pulls: transcript, content and key knowledge. |
A skill file from a public repository | You are installing code from GitHub. Read the file before you run it – a SKILL.md is instructions your agent will follow. Also: auto-captions get names, numbers and jargon wrong |
| 03 Give it eyes |
“Gemini reads YouTube natively, because Google owns both.” Key: free key from AI Studio. Bridge: YT Analysis MCP into Claude Code. Result: every word plus frames and sound. |
A Google API key, and an MCP server wired into your agent | That key is a credential, not a setting. Free tier still means a quota, and a key committed to a repo is a leaked key. See the caveats |
| 04 Build expert agents |
“Feed it the best videos on a skill and get an expert agent.” Framework: multi-skill – each video becomes a skill. Example: 10 editing tutorials in, editor agent out. |
Ten videos you chose, and a judgement about which ten | The whole thing now rests on “best”. The pipeline has no opinion about which ten videos are right – that choice is yours, and it is the entire product |
| 05 Verify not on the card |
— | One hour, and a question you already know the answer to | Without this the agent is confident in exactly the proportion the videos were confident, which is not the same as being right |
Why ten tutorials do not add up to one expert
Step 04 is the step everyone will act on, so it deserves the most care. “Ten editing tutorials in, editor agent out” is a lovely sentence, and the thing it quietly assumes is that ten tutorials on a subject form a coherent body of knowledge. They do not. Here is what you are actually feeding in.
They contradict each other. Ask ten editors about pacing, or about export settings, or about whether you should cut on the beat, and you will get at least three incompatible answers, all delivered with total certainty. An agent trained on all ten does not resolve the contradiction – it absorbs it and then picks one at random depending on how you phrase your question.
They are optimised for watch time, not for being right. A tutorial that says “it depends, here are four cases” loses to a tutorial that says “here is the one setting nobody tells you about”. The videos that rank are the confident ones. You are not sampling the field; you are sampling the field’s most confident ten per cent.
They age silently. A two-year-old tutorial about a menu that has since moved is not flagged as wrong anywhere – it just keeps teaching the old location. We have been round this on tool lists more than once: one card recommended a product that had closed over a year earlier. A tutorial corpus has the same problem with none of the visibility, because there is no footer saying when it was checked.
And an agent inherits tone, not accuracy. This is the part that catches people. If every source sounded sure, the agent sounds sure. Confidence is the easiest thing in the world to learn from a video and correctness is the hardest, so an agent built this way gets very good at the first one very fast.
None of that means do not build it. It means the build is not finished at step 04. It is finished when you have run step 05 and know where the agent is wrong.
Step 05: a verification pass you can run in an hour
You do not need a test suite. You need four questions, and you need to ask them before the agent does anything you care about.
| Check | How to run it | A failing answer looks like | What you do about it |
|---|---|---|---|
| 1. The known answer | Ask it something you already know cold from your own work | A fluent, confident answer that is subtly wrong – and no hedging anywhere in it | The corpus is bad or too narrow. Change the ten videos, not the prompt |
| 2. The contradiction | Ask it a question the ten videos genuinely disagree about | One confident answer with no mention that there is any debate | Tell the agent, in its instructions, to surface disagreement rather than resolve it |
| 3. The citation | “Which video said that, and roughly when?” | It cannot say, or it invents a plausible-sounding source | Require a source in every answer. An answer with no source is a guess with good grammar |
| 4. The date | “When was this published, and has anything changed since?” | It has no idea, because nothing in the pipeline recorded it | Record upload dates when you build the corpus. This is one line of work that saves you months |
Check three is the one with the best ratio of effort to value. If you require a citation on every answer, two things happen at once: you get a way to audit the agent, and the agent gets noticeably more careful, because answering “the third video, around eight minutes” is harder to fake than answering confidently.
This is the same discipline as any other list of claims. When we pulled apart a card of thirteen free alternatives, the useful work was not reading it – it was adding the column the card had no room for, and checking it. An agent built from ten videos is a card with sixty columns and no footer.
Four honest caveats
The free key is a credential. Step 03 puts a box on the diagram for the free API key from AI Studio. The card is right that it is free and right that it is the bridge. What a diagram cannot say is that an API key is a password. It goes in an environment variable or a config file your version control ignores – never in a committed file, never in a screenshot, never pasted into a chat. Keys get scraped out of public repositories automatically and within minutes. And “free” means a quota, not an absence of billing: if you wire a key to a paid project and run a hundred videos through it, somebody is paying for that.
A skill file is code. Step 02 installs a SKILL.md from a public repository. That file is a set of instructions your agent will follow, which makes it the same category of thing as a script you downloaded and ran. Open it. Read it. Check who maintains it and when they last touched it. This is not paranoia – it is the same two minutes you would spend on any dependency.
Watching and ingesting are different acts. Pulling transcripts and frames out of someone’s tutorial to build a product you use commercially is not the same as sitting and watching it, and the rules that cover it are not the rules of the YouTube player. Use the output to learn and to inform your own work; do not republish someone’s material, and do not treat a corpus of ten creators’ videos as your own content. If what comes out the other end resembles any one source closely, you have built a copy, not an agent.
The card’s footer is right and worth repeating. “Tools and features change often. Check each official source before you build.” This pipeline has four moving parts – a skill file, an MCP server, an API key policy, and a model – and all four can change independently. Anything in this post was true on 7 October 2026 and will need re-checking, exactly as the footer says.
What this is actually good for
With the fifth step added, the thing the card describes is genuinely strong – and it is strongest in a narrower place than “build an expert”.
It is excellent at extraction: pull the exact settings out of a forty-minute tutorial so you do not have to scrub through it; find the one timestamp where someone shows the thing you are stuck on; turn a rambling walkthrough into a checklist you can follow with your hands. That is a real saving every single week, and it does not require the agent to be right about anything – only to report accurately what the video said.
It is weaker at judgement: deciding which of three approaches is better, or what you should do with your channel. That is the part where ten confident videos produce one confident agent and nobody has checked anything. Keep that decision for yourself.
The split is the same one we keep arriving at from other directions – whether a tool captures, generates or reshapes your material tells you far more about what to trust it with than its feature list does. Extraction is reshaping. Expertise is not.
The one-line version
Four steps: the gap, the ears, the eyes, the agent. All four are right. Add a fifth – verify – and do it before the agent touches anything you would be embarrassed to publish.
And ask the cheap question before you turn on the expensive sense: does this video’s knowledge live in the words or in the pixels? Four in ten tutorials never needed the eyes at all.
Learning a lot does not make anyone good. Doing a lot does. That applies to agents too – an agent that has watched ten tutorials has learned a lot and done nothing, which is the exact condition this blog keeps warning people about, now automated.
If you want to be held to the doing part, we run a free challenge at BANI Academy. There is no barrier to entry – you commit a small fee that is refunded when you finish, and the only requirement is that you actually do the work for the length of the challenge. Build the agent if you want. Then go and cut a video with it.
If you have built one of these, what did it get confidently wrong first? Tell us in the comments under the original reel – the failures are more useful to everybody than the wins.
More on building a channel for a global audience at mmoyoutube.com.
3 Comments