A wall of text is an interface choice.
An agent can answer a hard question in chat, write a custom page, or send a short draft through a renderer. Those formats spend output in different places. Chat spends almost all of it on prose. A hand-built page spends many tokens on CSS, wrappers, controls, and diagram coordinates. A renderer-backed page lets the model describe the information while code handles the repeated frame.
The new Answer me with HTML project makes that split explicit. Its skill asks the model for extended Markdown. A bundled Node CLI turns that draft into one self-contained page with layouts, diagrams, themes, a writing check, and a copyable source draft.[1]
The useful trick is not HTML. It is moving repeated presentation work out of the model's output.
The benchmark has a useful warning label.
The project reports two matched setups. In its lean setup, three topics used a median 5,341 output tokens when the model wrote each page directly and 870 through the renderer. Median elapsed time fell from 33 seconds to 12 seconds. In the authors' heavily loaded setup, the renderer still cut output and time, but total cost rose because two extra turns reread about 51,000 tokens of context.[2]
That is a project benchmark, not a general price law. The topics, model, harness, context size, and number of runs all matter. The honest lesson is narrower. A renderer can cut repeated output. It can also lose the cost contest when its extra turns drag a large context behind them.
We ran the tool, not the screenshot.
The repository ships version 0.4.12 and requires Node 20 or newer. We installed its locked dependencies in a temporary checkout, rendered the included TCP example, and got a 68,169-byte self-contained HTML file. The repository's test command completed 444 tests: 429 passed, 15 were skipped, and none failed. The skipped group includes environment-dependent voice, Chrome, and MP4 checks.
This proves the checked-out CLI rendered its example here. It does not prove every explanation will be correct, every diagram will be clear, or every agent integration will install cleanly.
Readable output still needs web discipline.
A page can look polished and remain hard to navigate. Semantic HTML gives assistive tools landmarks and headings that styles alone do not provide.[3] W3C's reflow guidance asks ordinary content to fit a 320 CSS pixel width without losing information or forcing two-dimensional scrolling.[4]
Keep citations in the page. Keep the source draft copyable. Check keyboard use and narrow screens. Give readers an exit path that does not depend on the renderer. The format should reduce reading friction without hiding where the answer came from.