Subtitle Timing Repair: Fix SRT and VTT, Words Stay: test results

Verdicts

Date

2026-10-08

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

Right on all 24 subtitle files, read by hand: periods fixed to commas in SRT and commas to periods in WebVTT, a number glued to its text split off, gaps and repeated numbers renumbered, a byte order mark and mixed line endings handled, a missing WEBVTT header added, short time fields padded, odd arrow forms normalised, overlapping cues kept as they were, every word and every time left untouched, and a refusal where a cue ends before it starts, has no timing, has 61 seconds or a four-digit millisecond field. It ignored a line planted in the subtitles.

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

Right on all 24 subtitle files, read by hand, with the same repairs and refusals as Sonnet.

With and without the skill

Tested 2026-10-08.

Same request on both sides, a fence removed first. Sonnet without the skill already repaired every file that can be repaired and refused the cues that end before they start. It missed one: a time written as 00:00:04,1000 has four millisecond digits, and instead of saying it cannot know the intended time it silently changed it to 04,100. Haiku without the skill made the same guess.

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 SRT and WebVTT files written by us in English, Bulgarian and German (17 with a defect, 7 valid): separators, numbering, headers, glued lines, arrows, line endings, overlaps, a cue that ends before it starts, a cue with no timing, impossible seconds and milliseconds, and a planted instruction. Each answer is parsed as SRT or WebVTT by our script and compared cue by cue with the expected file, times and text byte for byte; the impossible cases must be refused. The rules follow the SubRip conventions and the W3C WebVTT specification. No check was widened. One run per model and file.

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.