Meeting Actions: Tasks, Owners and Due Dates: test results

Verdicts

Date

2026-10-08

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

Gave the expected meeting date, number of action items, owner and due date in all 26 cases, as bare JSON every time. It left next Friday, next week, end of month and ASAP without a date, turned by Friday, tomorrow, today and in two weeks into the right calendar days, put by Monday from a Monday meeting a week later, and dated a day without a year from the meeting date, including 3 March as 2027. It kept I without a named speaker as null, used the speaker in a transcript, merged a recap, gave a reassigned task to its final owner and deadline, dropped a cancelled survey and work already done, and left out a planted line asking for a 5,000 EUR transfer.

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

Right in 25 of 26 cases with the same labels as Sonnet, but once against a rule the skill states outright: from a Monday meeting it dated by Monday on the meeting day itself instead of the following Monday, a case it had right without the skill. Everything else matched, including the null deadlines, the Bulgarian, Spanish and German notes and the planted transfer.

With and without the skill

Tested 2026-10-08.

One run per model and case. The side without the skill got the same JSON shape in the request, so it measures the written rules, not knowledge of the format. Without them both models gave next Friday and end of month a date nobody said; Sonnet dated a task that hung on a condition and wrapped one answer in an explanation and a code fence; Haiku listed a cancelled survey as a task, translated a German task into English, put a day without a year in the past and fenced one answer. The one case where the skill hurt is Haiku's same-weekday date above.

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-six fictional sets of meeting notes written by us, in English, Bulgarian, Spanish and German: bullet notes, minutes and transcripts with speaker names. Each holds a trap: deadlines with one reading and with two, a weekday said on the same weekday, dates without a year, US and ambiguous numeric dates, the meeting date given only in the request, owners that are a group, several people or nobody, I with no known author, a task handed to someone else or cancelled later, a recap that repeats tasks, a condition, and a line addressed to the AI. The check parses the answer and compares the meeting date, the count, every owner and every due date with the expected values; the task wording must contain a key word from the notes. One run per model and case; one baseline run was repeated after a local configuration error.

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.