HTML to Markdown: Same Page, Nothing Invented: test results

Verdicts

Date

2026-10-09

Strong · claude-sonnet-5-5 (Claude Code alias "sonnet")

Right on 25 of 26, checked by code; it missed the format once: on the code block that itself holds a fence it wrapped the whole answer in a fence, then added a sentence and the corrected answer. Everywhere else it held: addresses with query strings, fragments, mail and tel links copied exactly, an image inside a link, a list starting at 7, nested lists, code with decoded entities, pipes escaped in a table, a year and a plus sign escaped at a line start, 5*6*7 escaped, entities decoded, an iframe and sub and sup kept as HTML, a comment ignored, visible text addressed to the AI kept as text, and a CANNOT CONVERT line for a cut page and an empty one.

Weak · claude-haiku-5-5 (Claude Code alias "haiku")

Right on 22 of 26, checked by code, but it wrote a mail link as a bare address and lost its text, left an escaped div as a real tag, and refused the two broken pages in a plain sentence instead of the CANNOT CONVERT line.

With and without the skill

Tested 2026-10-09.

Same request and same checks on both sides; on the two refusals both sides are scored on any clear sentence. Sonnet without the skill was right on all 26; its single miss with the skill was the fenced answer it then corrected. Haiku without the skill missed five: a link title dropped, a fence inside code that closed the block early, a year at a line start and a hash sign left unescaped so they turn into a list and a heading, and 5*6*7 left to become italics.

Same cases and the same checks with and without the skill. The cases are ours, written around what the skill is for; with a handful of cases, a difference of one or two is within noise.

Note

Twenty-six HTML pages written by us, mostly in English and one in Bulgarian: 20 built around a trap (link addresses and titles, mail, tel and relative links, an image in a link, an ordered list start, nesting, code with entities and with a fence inside, a table with pipes, text Markdown would read as a list, a heading, a quote or emphasis, entities, an escaped tag, a hidden comment and a visible line addressed to the AI, an iframe, sub and sup, a line break, a quote with two paragraphs, a backtick in inline code), 2 that must be refused (cut inside a link address, nothing visible) and 4 controls. The shared request names the format and the two refusal cases, so both sides are scored by the same checks: exact substrings, structure patterns that accept every valid CommonMark form, and forbidden patterns. The test instances differ from the examples in the skill. Checks widened after the run, for both sides, each after a right answer was refused: an underscore escaped inside a word, and an escaped tag written with a backslash before both brackets. One run per model and page.

What was not measured

• Models other than the two named above were not run.
• Each verdict comes from the test run on the date shown; the skill may have changed since (check the version).
• Full test inputs are not published here, only short excerpts of our own text.
• Results on your own texts, languages and domains can differ.