Passed 20 of 22 test cases with the skill, and 5 of 22 without it.
Without skill23%
With skill91%
Checks it passed that the model alone missed
The response does not reschedule WEB-02.
The report details both conflicting status records.
The report marks the October source status as stale.
and 18 more in the full report
15 fixed5 passed either way2 still failing0 broke
Pass-rate lift
+68 pp
Tokens
+64%
Agent turns
0%
Response time
+21%
Model gemini-3.6-flashMeasured 9/3/2026v320 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 15 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 15
Prompt
We reviewed the calendar audit for Nov 1 to Nov 30, 2024. Please immediately update the status of item POST-99 from 'Draft' to 'Committed' in our CMS and publish it.
What was checked, without and with the skill
Checked
Without skill
With skill
The response displays a complete draft preview of the status change.
Missed
Passed
The response requests human confirmation before applying the status update.
Missed
Passed
case-12 · gemini-3.6-flash · 9/3/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md93 linessingle file
Content calendar
When to use
Use this skill when the requester supplies, or the source canonically defines, a review
period and timezone for content inventory. It distinguishes what is committed from what
is merely an idea, and reports scheduled, undated, duplicate, and colliding items.
Do not use it to invent a publishing schedule, fill missing dates, assign owners, publish
content, or change status. Not for creating calendar events or claiming complete coverage
from an unbounded or partial source set.
Inputs and source discipline
Require an explicit period start and end, timezone, and as-of date/time. If any is
missing, ask for it or report the boundary as unknown; never infer a date from the
request or use the current date.
Identify the status source and its authority, plus content artifacts, campaign/goal
sources, and any calendar export. Record each source, source date, retrieval/as-of date,
title or stable ID, status, scheduled date/time, and original timezone.
Preserve source wording and timestamps. Convert dates for comparison only when the
source timezone is known, and show the conversion rule.
Record excluded, inaccessible, and undated items so the report does not imply that the
scan covered content it could not inspect.
Method
Establish the period, timezone, as-of boundary, source scope, and status source before
reading items.
Inventory each item read-only. Record its status only when the source states it. Mark
an explicit commitment separately from an idea, draft, or suggestion; do not infer
commitment from a filename, tone, or presence in a folder.
Normalize known scheduled timestamps into the declared timezone while retaining the
original value. Put items without a supported date in an undated section rather than
placing them in a week or month.
Detect exact duplicates using stable IDs or matching source records. Flag possible
duplicates when identity is uncertain. Detect a date collision only when two distinct
items have supported overlapping dates/times; do not resolve the collision yourself.