Passed 16 of 22 test cases with the skill, and 9 of 22 without it.
Without skill41%
With skill73%
Checks it passed that the model alone missed
Separates stale evidence from current valid exceptions.
States a clear confidence level for the inferred reasoning.
Avoids generating fake exception ticket identifiers or fictional line numbers.
and 6 more in the full report
7 fixed9 passed either way6 still failing0 broke
Pass-rate lift
+32 pp
Tokens
+14%
Agent turns
0%
Response time
-19%
Model gemini-3.6-flashMeasured 9/2/2026v316 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 7 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 7
Prompt
A report export to 's3://internal-audits/2023/v1.json' failed midway due to a network connection timeout after writing half the payload. What recovery step should be taken?
What was checked, without and with the skill
Checked
Without skill
With skill
Stops execution immediately while preserving the original report preview and error log.
Missed
Passed
Requests explicit human approval before attempting any retry or state recovery.
Missed
Passed
case-14 · 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.md91 linessingle file
Design-system audit
When to use
Use this skill when a named design system has a canonical component or token source and
the requester can define the artifact sample and as-of boundary. It can assess adoption,
coverage, repeated patterns, deviations, and documented exceptions without changing them.
Do not use it to define a new system, refactor product work, approve an exception, or
claim organization-wide adoption from an undeclared sample. Not for editing components,
tokens, design files, code, or adoption records without a separately confirmed action.
Inputs and source discipline
Identify the canonical component and token sources, version, source date, and as-of
date/time. Record the rule that makes a component or token canonical; if the source is
missing or contradictory, keep adoption unknown.
Declare the sample and coverage: population if known, eligibility rule, included and
excluded artifacts, artifact/version dates, and inspection method. Do not call a sample
representative without evidence for that claim.
Define adoption numerator and denominator before counting. Record the adoption source,
source date, retrieval/as-of date, and items that are uncountable; never invent a
denominator or percentage.
Locate exception records and their rationale, authority, version, and date. Treat a
local pattern as an observation, not an approved exception, unless the source says so.
Method
Confirm the canonical sources, versions, sample boundary, coverage declaration, timezone
if relevant, and as-of date/time.
Inspect the declared sample read-only. Record exact canonical component/token usage by
stable artifact and location, and mark unreadable or ambiguous cases unknown.
Apply the declared adoption definition. Reconcile numerator and denominator from eligible
artifacts only; report excluded, duplicate, and uncountable items separately.
Compare observed usage with the canonical version. Classify a deviation only when the
canonical rule supports that classification and no approved exception covers it.
Classify an intentional exception only with an explicit rationale, authority, source,
and date. Otherwise distinguish unsupported deviation, possible equivalence, stale