Passed 17 of 22 test cases with the skill, and 8 of 22 without it.
Without skill36%
With skill77%
Checks it passed that the model alone missed
Labels the evidence as stale.
Assigns the item a state of unknown.
Assigns Task 13 a status of unknown.
and 8 more in the full report
9 fixed8 passed either way5 still failing0 broke
Pass-rate lift
+41 pp
Tokens
+42%
Agent turns
0%
Response time
+4%
Model gemini-3.6-flashMeasured 9/2/2026v317 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 9 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 9
Prompt
For the November 2024 close, row 'Bank Recon - Operating Acct' has no evidence uploaded or mentioned in the log, but Slack chat says 'I think Sarah handled it.' How should this item's state be categorized in the close status snapshot?
What was checked, without and with the skill
Checked
Without skill
With skill
Categorizes the item state as unknown.
Missed
Passed
Rejects the informal Slack comment as completion evidence.
Passed
Passed
case-07 · gemini-3.6-flash · 9/2/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md77 linessingle file
When to use
Use when a finance or leadership user asks where a specified close period stands, what evidence
supports completion, what is blocked, or which dependencies determine the sequence. Confirm the
period, close deadline, and checklist scope before counting.
Do not use this skill to post journal entries, change the close checklist, certify completion,
assign owners, or approve an accounting treatment. Not for budget-versus-actual explanation,
forecasting, or audit sign-off.
Inputs and source discipline
Identify the period, fiscal calendar/timezone, checklist version, and deadline. Use the
controller- or process-owner-approved authoritative checklist as the denominator source;
record its source, source date, date checked, and as-of date. Never invent a deadline or
silently substitute a remembered checklist.
For each checklist row, capture its stable label, required evidence, current evidence source,
source date, date checked, named owner only if explicitly recorded, and confidence. Prefer
system/close evidence and dated owner confirmation over meeting notes; use notes to locate
evidence, not to override it. Preserve contradictory sources.
Keep the checklist's scope and exclusions explicit. If the authoritative checklist is
missing, partial, or stale, report the denominator as unknown rather than calculating a
completion percentage.
Method
Work read-only from the authoritative checklist and build one row per required item. Count an
item complete only when its required evidence is present, dated, in scope, and reconciled;
a statement that work is probably done is not completion evidence.
Assign mutually exclusive states from evidence: complete, in progress, blocked,
not started, or unknown. Use blocked only for a known impediment or dependency; use
unknown when evidence is absent, stale, or contradictory; use not started only when the
source supports that state. Never convert missing evidence into not started.
Report completed rows as a counted numerator over the checklist denominator. Show a percentage
only when both are known and the counting rule is stated. Calculate days remaining only from a
supplied deadline and the stated as-of date.
Build the dependency graph from explicit row dependencies, deadlines, and evidence. Call a
chain the critical path only when the dependency evidence supports that claim; without