The Drop Report · free, no account
Your agent read the words. It never saw the design.
Every web-to-markdown reader converts a page to prose and discards the rest. Paste a URL and see both halves: about what a reader carries away, and exactly what it leaves behind.
Try
What a reader keeps
Approximate. Readers disagree about nav, footers and alt text.
What it drops
Counted declarations from this page's CSS. Not a score.
Why nobody else can measure this
A markdown reader cannot tell you what it dropped, because it never parsed the stylesheet: the design layer was discarded before it had a name. A design tool cannot tell you either, because it never counts the words. Both halves have to come from one fetch of one page, which is what happens here.
The point is not that markdown is wrong. It is the right format for reading. The point is that an agent handed only markdown has the page's prose and none of its system, and will invent a visual system when asked to build something that matches. If it is going to write UI, give it a DESIGN.md as well as the text.
The same thing, as one call
curl "https://csscreme.com/api/decode-url?url=https://stripe.com"No key, no account, no credits, no multipliers. Add &pages=5 and it reads a sample of
the site instead of one page, which is the consistency scan. Agents can call
the same measurements through the MCP server as decode_url and
decode_site.
Questions
What is a Drop Report?
It is what a web-to-markdown reader throws away, measured. Markdown carries a page's prose and discards its design: colour, type, spacing, radii, corner and state. This reads both halves from a single fetch and tells you the size of each, so you can see what your pipeline is not carrying.
Why does this matter for a RAG pipeline or a coding agent?
Because an agent that only ever read the markdown of a page has no way to rebuild how it looked. It has the words and none of the system. If the agent is going to write UI, hand it the design layer as well as the text, or it will invent one.
Is the word count exactly what my reader returns?
No, and nothing here claims it is. Readers disagree about navigation, footers, boilerplate and image alt text, so this is honest as a magnitude and not as a guarantee. That is why every figure is written as "about". The design side is not an estimate: those are declarations counted in the CSS.
Does it run JavaScript?
No. It reads the served HTML and the stylesheets that HTML links to. When a page builds its content with JavaScript, the report says so in plain words rather than reporting a thin page as if it were the whole one. Stylesheets are static files, so the design measurements are usually complete even then.
Is there an API, and does it cost anything?
Yes and no, in that order. GET https://csscreme.com/api/decode-url?url=https://example.com returns the same JSON this page renders. There is no key, no account, no credit and no multiplier. The only limit is twelve requests a minute per address, which exists so one caller cannot spoil it for everyone.
Next: the full decoder · scan a whole site · what a crawler sees