One skill, ad images, pay-as-you-go routing

A seven-minute demo: one tool makes ad images, turns one into a video, pulls the colours out of it and codes a web page. The tool design is a well-built real thing. The conclusion in the title falls apart on a table the video itself puts on screen.

What the demo does — one line in, everything out

The human does one thing: “make me some ad images.” After that it runs:

1. Produces six images
2. Picks one and turns it into a video
3. Pulls the colours out of that image and codes a web page

The last step is the good one. Rather than a person choosing the colours, they are extracted from the artefact. So the image and the page match in mood automatically.

Cost was $0.73 for stage one and about $1.16 for stage two.

The good design — “where in the flow does the budget check sit”

The most instructive part.

When you give an AI a job that spends money, checking the budget after the call is too late. It has already gone.

This tool adds up unit price × count before calling and checks whether it exceeds. If it does, it reduces the count or switches to a cheaper model.

When you check	Result
After the call	Money already spent
Before the call	You can stop it going out

Another one — “the tool pre-empts the most common misdiagnosis”

Small and good.

The error code returned when your balance is zero is the same one returned when your credentials are wrong. So people spend hours looking in the wrong place — is my key bad?

This tool, on that error, says “this is not a key problem, it is a balance problem.”

The most common misdiagnosis, pre-empted by the tool.

And what I actually took — “state quality in one word”

I took one thing. It cost nothing, so it went straight in.

When you ask an AI for quality, “do it well” has no effect. Instead, write this:

What to write instead
“Treat this as a showcase piece”
“Don’t fake art with gradients — use a real image”
“Pull the colours from the result”

The second is especially good. It names the cheap trick an AI reaches for when it is phoning it in. Naming the trick to avoid is far stronger than please do a good job.

But the title’s premise is refuted by the video’s own table

The subject of the piece.

The title is “I cancelled my subscription” — the claim being that paying per use beats a flat subscription.

And in a comparison table the video puts on screen, for video generation the subscription was cheaper than all three pay-as-you-go options.

The presenter concedes it near the end.

So the title says cancel and the video contains the case for not cancelling.

And something heavier

The motivating reason for the video was a terms-of-service problem with a particular service.

The terms text they put on screen says the opposite.

So it was put on screen and not read. And that was the starting point for the whole video.

→ What I took: putting evidence on screen and that evidence supporting your claim are different things. Surprisingly few people read what is shown.

Reproducibility, stated first

The limits of this source, written down.

Item	State
Cost measurement	One of five platforms verified against an actual invoice. The rest are computed from public price lists
The tool’s code	Not in a public repository — shared conditionally

Because of the second, I cannot try it. So everything here reaches as far as what was visible on screen.

The rest of the verdict

What	Verdict	Why
Image → video → colour extraction as a unified look	Parked	Revisit after the deployment decision
Pay-as-you-go image APIs	Parked	Revisit when I actually generate assets in bulk. ⚠ The cheapest of them is off the list until its terms question resolves
Budget ceiling and retry	Rejected	I am on a flat rate, so no cost pressure, and retry is already handled elsewhere

And two things I was already doing came out of the takeaway count — my video-analysis tool already uses pay-as-you-go, and listing trigger conditions in a tool’s description was already standard for me.

Counting “newly learned” without subtracting what you already do inflates the number.

The detailed record starts here

The skill design itself is a well-built real thing — provider routing, a budget ceiling and output chaining are baked into the procedure, and the user writes not one line of image prompt. The problem is the title. The basis for “I cancelled my subscription” collapses on a table the video itself puts on screen — for video generation the subscription beats all three pay-as-you-go providers. The presenter concedes it near the end. And there is a heavier one. The stated reason this video exists contradicts the very screenshot it displays. What comes across is not the cost conclusion but one prompt-design technique.

What it is

A seven-minute tutorial showing one skill produce several ad images, turn one of them into video, pull colours out of that image, and code a one-page site — as a single chain. The claim is that the generation APIs behind it should be routed to three pay-as-you-go providers instead of a subscription.

The user writes not one line of image prompt. Give it “what (the product) + a reference (images) + constraints (rules)” and the skill handles concept planning, model choice, prompt writing and the API calls.

Step 1, six images	$0.73 against a $2 budget
Step 2, site plus video	~$1.16 against $2
Final artifact	a 1.1MB single HTML file with a five-second muted loop as background
Measurement scope	⚠ only one of five platforms has a real invoice. The rest is arithmetic on public rate cards
Reproducibility	⚠ low. The skill code isn’t in a public repository; it’s shared conditionally

The design decisions

Name	What it does
Trigger words piled into the blurb	Normally only the blurb is in view; when a request matches, the whole body gets read. So stuffing trigger words into the blurb is the trick
Procedure and calls, separated	The top document holds procedure and model criteria; the actual call syntax lives in per-model files. New models copy a template
Budget is part of the procedure	Unit price × count is summed before any call. Over budget, it cuts shots or drops to a cheaper model
Key lookup order + a prohibition	env var → folder file → ask the user. Never store a key in the skill file — it’s a shared object
Pre-empting a misdiagnosis	The 403 you get at zero balance is explained as “not a bad key, an empty account.” The most common misread, headed off by the skill
Quality bar in one word	“treat it as a showcase piece” · “don’t fake art with gradients, use real images” · pull the palette out of the result

What broke — four things

❶ The title’s premise is refuted by the video’s own table.

❷ The heaviest one — the stated motive contradicts the evidence on screen. The narration says “even your private content may be used for AI training”, but the official post displayed on screen says precisely the opposite — private stays private, only material posted to public pages is used. The reason this video exists conflicts with its own screenshot.

❸ The cheapest figure is one the presenter doesn’t trust. “I think it doesn’t go through the API but rides an actual subscription. I’ll look into it and share.” ⚠ That reservation isn’t small — if true, it means a terms-of-service risk is hidden inside the price. Layered on top is his own admission that “the servers are unstable.”

❹ One real invoice, four rate-card calculations. And the decisive number is not the measured one.

Held against my own setup

First I struck out what was already handled. My video analysis tool already uses a pay-as-you-go transcription API, and piling trigger words into a blurb is already standard across several of my skills. Neither counts as something this video teaches me.

Technique	Where it would go	Verdict
Quality bar declared in one word	report generation in the video tool · writing prompts	adopt — zero cost, immediate
Image → video background + palette extraction	this site’s landing / the game page	hold. Reopens after the domain and deploy decision
Pay-as-you-go image and video APIs	game asset generation	hold. Reopens when assets are actually generated in bulk. ⚠ The cheapest provider is out until the ToS reservation clears
Budget ceiling + single retry	—	rejected. Flat-rate, so no cost pressure, and retries are already a pipeline gate’s job
Per-model recipes + template	—	rejected. There is no multi-provider routing to do in the first place
Gallery + spend log file	—	rejected. The library index and the run-folder convention already do this

Verdict

One adoption — and even that one is still an estimate.

⚠ This video has no control. It never shows the run with “showcase” removed. The output is genuinely good, but whether that came from the word, the four reference images, or the model was never separated. With a sample size of one there is no confidence interval to give.

So it is adopted and labelled [estimated], with the promotion path written down — hold back 10% of future reports built without that sentence, and compare. That is when it becomes [measured].

And the weakest link in this teardown, stated plainly. ⚠ I never saw the skill code. Every structural claim was reconstructed by reading the presenter’s screen out of the frames, and whether the real file behaves as the screen says was not verified. The verdicts above are about “is this a good design”, not “does this implementation work.”

Three things I didn’t do — ① I never checked the terms of service directly; I read the screenshot they displayed. Anyone citing refutation ❷ has to start from the original. ② What the cheapest provider actually rides can’t be confirmed. ③ ⚠ The five rate cards are a snapshot. Image generation prices move quarterly, so do not cite them at adoption time without re-checking.

The lesson I paid for here is: read the tables to the end of the same talk. The first table makes the title right and the second makes it wrong. The presenter honestly concedes it near the end, and yet the title and thumbnail were built from the first table.

And ⚠ a good design and a good conclusion do not arrive as one package. The skill structure has things worth taking; the cost conclusion is unusable. Accept them as a bundle and you buy a wrong conclusion along with the design.