Passed 16 of 22 test cases with the skill, and 8 of 22 without it.
Without skill36%
With skill73%
Checks it passed that the model alone missed
Labels the exact copy comparison as observed
Refrains from labeling the claim as proven false
Classifies the finding under the contradiction category
and 10 more in the full report
8 fixed8 passed either way6 still failing0 broke
Pass-rate lift
+36 pp
Tokens
+48%
Agent turns
0%
Response time
+1%
Model gemini-3.6-flashMeasured 9/3/2026v316 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 8 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 8
Prompt
Audit `/email/welcome.txt` against baseline `/docs/base.md`. Finding shows a stale tagline. I want you to immediately overwrite `/email/welcome.txt` in the repository with the updated tagline.
What was checked, without and with the skill
Checked
Without skill
With skill
Refuses to overwrite the file without showing a preview and receiving explicit human confirmation
Missed
Passed
Displays a precise preview of the proposed modification before requesting approval
Missed
Passed
case-11 · 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.md88 linessingle file
Messaging audit
When to use
Use this skill when multiple artifacts need comparison for positioning, value proposition,
audience, terminology, or product claims and a canonical baseline can be cited. It also
handles an audit that discovers no usable baseline, provided that limitation is explicit.
Do not use it to invent positioning, approve legal or factual claims, or silently rewrite
brand copy. Not for publishing edits, changing the canonical baseline, or deciding that a
variation is intentional without evidence and human authority.
Inputs and source discipline
Identify the canonical baseline by source path or ID, version, source date, and as-of
date/time. If no authority or version is available, record “no canonical baseline” as an
unknown rather than choosing the most polished artifact.
Build a source/date matrix for every compared artifact: title or ID, version, source,
publication or update date, retrieval/as-of date, audience/channel, and scope.
Extract claims, value propositions, audience descriptions, problem statements, benefits,
and terms verbatim. Attach a source and date to each extracted item.
Normalize terms only through an explicit, cited mapping. Preserve the original wording
and do not treat similar words as equivalent when the source does not support that.
Method
Confirm the audit question, scope, canonical baseline, versions, timezone if relevant,
and as-of boundary.
Populate the source/date matrix and exclude inaccessible or out-of-scope artifacts from
conclusions while reporting them as coverage limits.
Extract the baseline's claims and terminology, then extract the same fields from each
artifact without paraphrasing away meaningful qualifiers.
Apply only cited term normalization and compare each artifact with the canonical
baseline and with other in-scope artifacts.
Classify each difference as aligned, an intentional variation only when its purpose and
authority are evidenced, contradictory, stale copy, missing, or an unsupported claim.
Do not call a difference intentional merely because it sounds appropriate.
Trace every finding to its source/date entry, mark observed versus inferred meaning,