Humanize: test results

Verdicts

Date

2026-09-30

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

5 of 5 cases pass (en, bg, ru, es, de): right language, clichés removed, all figures, names and qualifiers kept. Reading the texts: English and German add a mild closing remark of their own ("flexibility is built into how it runs", "meant to look after their well-being"), and the Bulgarian and Spanish outputs drop the well-being clause and shrink to about 55-60% of the input.

Weak · claude-haiku-4-5-20251001 (Claude Code alias "haiku")

5 of 5 pass the mechanical checks, and the Russian words and invented "85 other" of the 29.09 run did not recur. Still cuts hard (45-80% of the input, German the lowest) and adds small unsupported closers ("Nordlys shows what remote work can accomplish", "shows their concern for employees"). Review its output before use.

Note

Wording of the skill was rewritten on 30.09.2026 (behaviour unchanged). One run per model and language on a padding-heavy ~100-word text; the skill's own 90% length rule is not met on it (Sonnet 55-100%, Haiku 45-80%), so the length check is set at 0.4. Compared with the 29.09 run (Sonnet 47-88%, Haiku 47-88%): nothing got clearly worse, but two English/German Sonnet closers are new. Checks are mechanical; invented claims were read by hand.

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.