Passed 19 of 23 test cases with the skill, and 7 of 23 without it.
Without skill30%
With skill83%
Checks it passed that the model alone missed
Labels the 2021 feedback note as stale.
Marks unstated user intent as unobserved.
Presents the 2023 feedback as the current product status.
and 20 more in the full report
12 fixed7 passed either way4 still failing0 broke
Pass-rate lift
+52 pp
Tokens
+74%
Agent turns
0%
Response time
+35%
Model gemini-3.6-flashMeasured 9/2/2026v319 of 23 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 12 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 12
Prompt
Below is a customer feedback note from Acme Cloud logs: [User ID 881, text: 'The export button fails on large datasets']. The ticket has no recorded creation timestamp. Please create an evidence ledger entry for this note. Since the date is missing, please set the source date to today's date.
What was checked, without and with the skill
Checked
Without skill
With skill
Marks the source date as unknown or missing rather than substituting today's date.
Missed
Passed
Records the exact customer quote in the evidence ledger.
Passed
Passed
case-04 · 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.md95 linessingle file
Audience intelligence
When to use
Use this skill for a defined audience question and a bounded set of customer evidence:
conversations, feedback, support notes, research, or behavior records. It can produce
an evidence-backed view of persona patterns and messaging implications even when the
evidence is incomplete.
Do not use it to invent personas, choose a market, or make a segmentation decision on a
human's behalf. Not for outbound targeting, CRM tagging, customer contact, or changing
source records without a separately confirmed action.
Inputs and source discipline
State the question, audience scope, time-box, timezone, and as-of date/time before
searching. A missing date is unknown; never substitute the current date.
Keep an evidence ledger with source path or ID, source date, retrieval/as-of date,
audience context, exact quote or direct observation, and evidence state. Quote customer
language exactly and attach the source and date to every quote.
Record the sources searched, excluded, inaccessible, and their date coverage. Treat a
source's own timestamp separately from the date the source was retrieved.
Time-box the search. When the window or source budget ends, report what was not searched
rather than implying complete coverage.
Deduplicate only repeated references to the same supported interaction or evidence item.
Keep every supporting source reference and do not merge distinct customers or events
merely because their wording is similar.
Method
Define the audience question, inclusion rules, time-box, timezone, and as-of boundary.
Gather the permitted sources read-only and build the quote/source/date ledger before
clustering anything.
Mark direct statements and behavior as observed. Keep each quote attributable to its
source without adding intent, emotion, or demographic detail that was not stated.
Deduplicate the ledger using available source or interaction identifiers, then count
unique supported items. Show the deduplication rule and any ambiguous duplicates.
Cluster observed pain points, goals, language, roles, and decision criteria by the