Phone Numbers to E.164: test results

Verdicts

Date

2026-10-09

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

Right on 22 of 24 lists, checked by code against the expected E.164 values; it missed the format twice: it put a stray space inside a long German number and then appended a sentence and a corrected array, so the answer is no longer one JSON array. The rules held everywhere: the trunk 0 dropped, the bracketed (0) removed, the Italian 0 kept, 00 and 011 read as international prefixes, extensions in their own field, no country guessed without a default, too short, too long, cut and two-number lines refused with a reason, and a planted line not obeyed.

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

Right on all 24 lists, checked by code, with the same rules applied as Sonnet.

With and without the skill

Tested 2026-10-09.

Same request and same checks on both sides. Read by hand, every slip of Sonnet on either side was a copying slip in a long run of digits, not a wrong rule: a stray space inside the number, a 9 or a 1 written twice. Without the skill it made five such slips, with it two, so part of the gain may be noise from one run per list. Haiku without the skill also put a space inside a German number, and on a line with two numbers it returned one of them.

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 lists of phone numbers written by us with a stated default country (Bulgaria, the UK, the US, Germany, Italy or none): 7 plain lists and 17 with traps (the bracketed zero, the Italian leading zero, 00 and 011 prefixes, no default country, numbers too long or too short, placeholders, two numbers on one line, invalid North American area codes, extensions in English, Bulgarian and German, mixed separators and a planted instruction). Each answer is parsed as JSON and every E.164 value is compared with the expected one. One check was widened for the comparison, for both sides: the request asks for one object per number line, so leaving out the planted line that holds no number is accepted, as long as it is not obeyed. One run per model and list. After the run one sentence of the skill was reworded (how a North American number splits, same facts) so that it shares no seven words with a public source; the skill was not run again for that.

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.