Trusting the Output

How to catch the confident-wrong before it ships.

An agent sounds the same when it is right and when it is wrong. The skill is not spotting a bad tone; it is checking the claim.

Craft guidanceShortcut editorial, reviewed 9 Oct 2026

In 30 seconds

Ask for the source, then open it
Requirement, recommendation and convention are different things
Check numbers and names first
Old information is stated as current
Decide how wrong you can afford to be

What goes wrong

Hallucinated UX standards
Fake citations
Fabricated WCAG requirements
Outdated product information
Incorrect design-system values
Invented features
Generic recommendations presented as rules
Research over-generalised
Accessibility claims never tested
Unsupported UI patterns

Trust check

What is the source?
Is it current?
Is this a requirement, a recommendation or a convention?
Did the AI infer something?
Can I verify it?
Does it match the product context?
What happens if this is wrong?

Label what you keep

Verified: You checked it against the original source.
Needs review: Plausible, not yet checked.
Subjective: A judgement. Reasonable people would differ.
Unsourced: No source given, or the source could not be found.
Outdated: True once. Check the current version.

What kind of claim is it?

Strongest first. An agent will state all six in the same voice.

Standard: A published requirement, such as a WCAG success criterion.
Platform guidance: What Apple or Google recommend for their own platforms.
Design system: A rule inside one organisation's system.
Industry convention: What many products do. Nobody requires it.
Craft recommendation: A senior designer's judgement.
Preference: Taste.

Confident-wrong, worked through

The AI says
Needs review
"WCAG requires every touch target to be 44 × 44 px."
It mixes three sources and attaches Apple's number to WCAG.
WCAG 2.2, Level AA
At least 24 by 24 CSS pixels, with exceptions.
Standard
WCAG 2.2, Level AAA
44 by 44 CSS pixels. AAA is rarely the level a product is held to.
Standard
Apple
A hit region of at least 44 × 44 pt.
Platform guidance
Android
Touch targets of at least 48 × 48 dp.
Platform guidance
The number is real. The attribution is wrong, and the units are not interchangeable.
Sources for this example
Sources
WCAG
WCAG 2.2, SC 2.5.8 Target Size (Minimum)(opens in a new tab)
Authority
Standard requirement
Published
12 Dec 2024
Last verified
Verified 7 Oct 2026
WCAG
WCAG 2.2, SC 2.5.5 Target Size (Enhanced)(opens in a new tab)
Authority
Standard requirement
Published
12 Dec 2024
Last verified
Verified 7 Oct 2026
Apple
Apple HIG, Buttons(opens in a new tab)
Authority
Design-system convention
Last verified
Verified 7 Oct 2026
Material
Android Accessibility Help, Touch target size (cites Material Design)(opens in a new tab)
Authority
Design-system convention
Last verified
Verified 7 Oct 2026
The AI says
Subjective
"Card padding should always be 24 px."
It presents a common habit as a rule.
Shortcut
16–24 px is a common desktop range. No standard sets card padding.
Industry convention
A reasonable starting point stated as a requirement. Treat it as convention.
The AI says
Unsourced
"Use color.surface.raised.subtle from your design system."
The name looks right and may not exist. Agents complete patterns.
Your design system
Only the tokens in your own file are real. Check the name against it.
Design system
Plausible is not the same as present. Search your tokens before using it.
The AI says
Unsourced
"Research shows users prefer bottom navigation."
No study, no users, no context. It reads as evidence and is not.
Ask the agent
Which study, which users, which year? If it cannot say, there is no finding.
Unsourced
A generic statement dressed as research. Do not carry it into a deck.

Go deeper in Shortcut

What kind of claim is it?

Strongest first. An agent will state all six in the same voice.

Standard

A published requirement, such as a WCAG success criterion.

Platform guidance

What Apple or Google recommend for their own platforms.

Design system

A rule inside one organisation's system.

Industry convention

What many products do. Nobody requires it.

Craft recommendation

A senior designer's judgement.

Preference

Taste.

Confident-wrong, worked through

The AI says

Needs review

"WCAG requires every touch target to be 44 × 44 px."

It mixes three sources and attaches Apple's number to WCAG.

WCAG 2.2, Level AA: At least 24 by 24 CSS pixels, with exceptions.; Standard
WCAG 2.2, Level AAA: 44 by 44 CSS pixels. AAA is rarely the level a product is held to.; Standard
Apple: A hit region of at least 44 × 44 pt.; Platform guidance
Android: Touch targets of at least 48 × 48 dp.; Platform guidance

The number is real. The attribution is wrong, and the units are not interchangeable.

Sources for this example

Sources

WCAG
WCAG 2.2, SC 2.5.8 Target Size (Minimum)(opens in a new tab)
Authority
Standard requirement
Published
12 Dec 2024
Last verified
Verified 7 Oct 2026
WCAG
WCAG 2.2, SC 2.5.5 Target Size (Enhanced)(opens in a new tab)
Authority
Standard requirement
Published
12 Dec 2024
Last verified
Verified 7 Oct 2026
Apple
Apple HIG, Buttons(opens in a new tab)
Authority
Design-system convention
Last verified
Verified 7 Oct 2026
Material
Android Accessibility Help, Touch target size (cites Material Design)(opens in a new tab)
Authority
Design-system convention
Last verified
Verified 7 Oct 2026

The AI says

Subjective

"Card padding should always be 24 px."

It presents a common habit as a rule.

Shortcut: 16–24 px is a common desktop range. No standard sets card padding.; Industry convention

A reasonable starting point stated as a requirement. Treat it as convention.

The AI says

Unsourced

"Use color.surface.raised.subtle from your design system."

The name looks right and may not exist. Agents complete patterns.

Your design system: Only the tokens in your own file are real. Check the name against it.; Design system

Plausible is not the same as present. Search your tokens before using it.

The AI says

Unsourced

"Research shows users prefer bottom navigation."

No study, no users, no context. It reads as evidence and is not.

Ask the agent: Which study, which users, which year? If it cannot say, there is no finding.; Unsourced

A generic statement dressed as research. Do not carry it into a deck.