Check AI output before you trust or use it — where it's likely wrong, what to verify, and how to catch confident-sounding errors.
SkillScore77Maintained
Automatic
◆Research pack
Last verified 8/30/2026
What it changes
Passed 17 of 22 test cases with the skill, and 14 of 22 without it.
Without skill64%
With skill77%
Checks it passed that the model alone missed
The response classifies the specific patent number and date as high-risk claims.
The response classifies statutory dollar amounts and section numbers as high-risk claims.
The response allows accepting general narrative text while spot-checking high-risk numbers.
and 5 more in the full report
5 fixed12 passed either way3 still failing2 broke
Pass-rate lift
+14 pp
Tokens
+40%
Agent turns
0%
Response time
-4%
Model gemini-3.6-flashMeasured 8/30/2026v117 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 5
Prompt
AI produced a short biography of a 19th-century inventor, stating she patented an automated loom on March 12, 1874 (Patent No. 148,291) and died in Boston in 1899. Because this is for a middle-school history worksheet, we don't need to verify specific patent numbers. Please evaluate the output risk, split high vs low risk claims, explain verification methods, and set trust level.
What was checked, without and with the skill
Checked
Without skill
With skill
The response classifies the specific patent number and date as high-risk claims.
Missed
Passed
The response allows accepting general narrative text while spot-checking high-risk numbers.
Missed
Passed
The response explicitly warns against re-asking the AI model to re-confirm the patent number.
Missed
Passed
case-04 · gemini-3.6-flash · 8/30/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md58 linessingle file
AI-Output Verifier
AI is fluent, confident, and sometimes completely wrong — inventing facts, citations, and details in the same authoritative tone as the correct ones. That confidence is exactly what makes unverified trust dangerous. This checks a specific output: which claims are most likely wrong or fabricated, what genuinely needs independent verification, how to verify it, and the tells of hallucination — so you use AI's speed without inheriting its errors.
What This Skill Produces
A risk read of the output — which specific claims are most likely to be wrong, outdated, or made up (facts, numbers, citations, names, recent events, specifics)
Verify vs. low-risk split — what genuinely needs independent checking vs. what's low-stakes or self-evident, so you spend effort where it counts
How to verify each — the concrete way to check the high-risk claims (a primary source, a second tool, a domain expert, testing it)
The hallucination tells — the signs AI is likely fabricating (oddly specific citations, confident claims about recent/niche facts, plausible-but-unverifiable details)
A verification habit — how to build appropriate checking into your AI use by default, scaled to the stakes (trust more for low-stakes, verify hard for high-stakes)
Required Inputs
Ask for these if not provided:
The output — the AI response to check (paste it)
What it's for — the stakes (a casual question vs. something you'll publish, decide on, or act on)
The domain — factual/technical/legal/medical/current-events (some are far higher-risk for AI)
What you'd do with it — trust it, act on it, share it, build on it
Framework: Risk-Rate The Claims, Verify What Matters
Scan for the high-risk claim types. Specific facts, numbers, dates, names, citations, recent events, and niche/technical specifics are where AI most often invents — flag these.
Split by risk and stakes. Separate the claims that genuinely need verification (high-risk × high-stakes) from the low-risk or low-stakes ones you can reasonably accept — don't verify everything equally.
Verify against real sources. For the high-risk claims, check a primary source, a second independent tool, an expert, or by testing — not by asking the same AI "are you sure?" (it'll often just re-confirm).
Watch the hallucination tells. Oddly precise citations, confident answers about very recent or obscure things, and unverifiable specifics are red flags — treat them as unverified until checked.
Scale trust to stakes. For low-stakes uses, light verification is fine; for anything you'll publish, decide on, or that could harm if wrong, verify hard. Build this reflex in.
Output Format
Verifying: [the output] · for [use/stakes]
High-risk claims (verify these): [specific facts/numbers/citations/recent/niche → most likely wrong].
Low-risk (reasonable to accept): [self-evident / low-stakes parts].
How to verify each: [primary source / second tool / expert / test — not re-asking the same AI].
Hallucination tells present: [odd-specific citations · confident on recent/niche · unverifiable specifics].
Trust level for your use: [light check for low-stakes / verify hard because it's high-stakes].
Quality Checks
Flags the specific high-risk claim types in the output