Markdown Table Repair: Fix the Table, Keep Every Cell: test results

Verdicts

Date

2026-10-08

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

Right on all 21 setups, read by hand: it added the missing alignment row, escaped the pipe inside inline code and inside an HTML cell, kept 007, 1.10, 1,250, the double spaces and the backslashes in a path, turned tab-separated text into a table, padded short rows at the end, ignored a planted cell that told it to reply OK, and answered the fixed CANNOT REPAIR line for a row with one cell too many (English and Bulgarian), for an alignment row with two colons over three columns, and for a sentence that only mentions a pipe.

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

Right on 20 of 21 by the skill's own checks, but it missed one detail on the tab-separated text: it turned the short last line into a table row with a fourth, empty cell, so that row is wider than the header. The cell text was right and every other setup matched, including the refusals for the long row, the unmatched alignment row and the sentence with a pipe.

With and without the skill

Tested 2026-10-08.

Same request on both sides; the bare side is scored on the table in any phrasing, fence removed first, and a clear refusal in any words is accepted where the skill refuses. Read by hand, Sonnet without the skill did the 17 plain repairs alone: missing alignment row, pipes in code and HTML cells, numbers and spaces untouched, tabs, short rows, the planted order. It missed the four places where the honest answer is a refusal: it invented a column for a row with one cell too many (English and Bulgarian), guessed a third alignment for two colons over three columns, and built a table out of a sentence. Haiku without the skill missed the same four, merging the extra cell into its neighbour, adding an empty header, guessing the alignment. Haiku with the skill still added an empty fourth cell to a tab-separated line.

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-one inputs written by us (15 with a trap, 6 controls that are already valid or only need the frame): a pipe inside code, a long row, a missing alignment row, a short row, tabs, an HTML table, a Bulgarian, a German and a Spanish table, a planted instruction in a cell. Each answer is scored by one regex built from the expected cells, byte for byte; spaces around the pipes, edge pipes, the number of hyphens and missing empty cells at the end of a short row are free, a code fence, a sentence around the table, a moved or changed cell or an unescaped pipe fails. Rules re-checked on 2026-10-08 against the GitHub Flavored Markdown tables extension (read-only fetch); the refusal rules are our own decision and not a claim about the renderer. Checks widened after the run: for the alignment row with two colons over three columns, a table without any colons now passes next to the refusal, on both sides, because dropping the colons guesses nothing; a table that keeps or invents a colon still fails. One run per model and case. The sets of the with and without side use the same inputs and the same request.

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.