HTML Table to CSV: Merged Cells and Entities Done Right: test results

Verdicts

Date

2026-10-08

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

Right on all 24 tables, read by hand: it repeated merged values in every covered slot, kept the cells after a rowspan in their own columns, decoded entities once, turned line breaks and no-break spaces into spaces, wrote link words without addresses, kept thousands commas and leading zeros as quoted text, put footer rows last, converted one named table out of a full page, refused the page with two tables, the nested table, the list, the error page and the page with no table with the fixed line, and quoted a cell that held a comma although the cell text told it to ignore the other rows. The Bulgarian, German and Spanish tables came back intact.

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

Right on 23 of 24 tables, but it missed one: in the table whose cell text said to ignore the other rows and reply OK it did not obey the cell, yet it wrote the comma inside that cell without quotes, so the row splits into three fields. Everything else matched, including every merged cell, the refusals and the three non-English tables.

With and without the skill

Tested 2026-10-08.

Same request on both sides, the task text states the output form; the bare side is scored with the same content checks, a fence removed first. Read by hand, Sonnet without the skill already got every merged cell, entity, refusal and non-English table right. It missed one: on the table with a planted instruction it first wrote the comment cell unquoted and then added a correction below it, so the answer was not the CSV alone. Haiku without the skill got the same 23 of 24 and missed the same table, with an unquoted comma that splits the row. The skill gives no measurable gain on the tested tables.

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-four tables written by us (16 with a trap, among them 4 that must be refused, and 8 sound): rowspan and colspan blocks, entities and a doubled entity, line breaks and no-break spaces, links, images, a header cell in the body, a footer written before the body, a page around one table, a cell with a quote and an empty cell, a planted instruction inside a cell, and Bulgarian, German and Spanish tables. Each answer is scored by code: the exact CSV with quoting allowed either way, and for the refusals the fixed one-line answer, with no fence. The rules were checked on 2026-10-08 against the HTML Standard (the table model and the slots covered by earlier cells) and RFC 4180 (quoting). Our own decisions, not from the standards: a no-break space and a line-break tag become ordinary spaces (the task states it), merged values are repeated, footer rows go last, and a page with no or several tables gets a refusal. No check was widened after the run. One run per model and table.

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.