Fewer tokens

The agent writes data, not a web page: the package owns the layout, the styles, the controls, and the code that sends answers back. For the four example pages below, writing the blocks took 1,253 output tokens; asking the same model for one self-contained HTML page with the same content and working controls took 12,364, about 10 times as many (90% saved).

Page	Blocks	HTML page (median)	HTML runs	Saved
Decision: alert, options, form	347	2,770	1,651 – 2,917	87%
Status: stats, steps, table	412	3,726	2,145 – 3,895	89%
Structure: tree, mermaid, diff	338	3,719	3,514 – 4,064	91%
Review: review, suggest	156	2,149	2,008 – 2,478	93%
Total	1,253	12,364		90%

The saving repeats on every revision: the agent changes a few YAML lines instead of rewriting markup and script.

How it was measured

Counts are Claude output tokens from claude -p --model claude-sonnet-5-5 (3 to 5 HTML runs per page; the block count is the file echoed back verbatim), measured on 2026-11-10 with scripts/token-bench.sh. None of the HTML pages had the feedback panel for commenting on a selected passage, which show_doc adds to every page.

The four pages