Teardownspublic

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

The skill design is a well-built real thing. The conclusion in the title, though, falls apart on a table the video itself puts on screen.

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.

structure — procedure on top, call recipes underneath one orchestration doc procedure · model criteria trigger words piled into the blurb one file per model an "API recipe" — calls only new models copy a template key rules — written as if the skill gets shared env var → folder file → ask the user never route to a provider with no key five steps — the budget check sits ahead of any call parse request plan the shots route · compute budget prompt · generate file it · log the spend unit price × count is summed before calling — over budget, it cuts shots or drops to a cheaper model a failure gets exactly one retry — a ceiling put there to protect the budget
The point of this design isn't the model list — it's which step the budget check sits in.

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.

per image — read this far and "cancel the subscription" holds payg A $0.05 ⚠ a figure he doesn't trust either subscription $0.31 payg B $0.38 per video — the conclusion reverses inside the same talk subscription $1.23 — the cheapest here payg A $1.32 payg B $1.94 ⇒ what survives is the much narrower "a few images now and then, go pay-as-you-go" ⚠ the subscription side treats the monthly fee as sunk and counts credits only — valid only if you burn it
Two tables in one talk point opposite ways. The title was made from the first of them.

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.